fix(compilers/openapi): keep undeclared keys on nested objects

[OmarAlJarrah]

This stacks on two open PRs and must merge after both. Its base is
fix/openapi-extensions-everywhere (#345), which adds the Unmodeled carriers
the objects below need, and it merges fix/openapi-unknown-key-census (#356),
which adds the census reader itself. #356 is a sibling of #345 rather than a
parent, so its commit appears in this branch's history; once both land the diff
here collapses to the last commit.

Summary

A key the OpenAPI model names no field for was kept at the objects that lower to
a node with an Unmodeled map of their own, and dropped in silence at the
objects nested inside one. Those had nowhere to land until the carriers went in.
This reads the census at each of them:

object	carrier	entry key
components	document	openapi:components/definitions
tag externalDocs	document	openapi:tags/0/externalDocs/title
operation externalDocs	operation	openapi:externalDocs/title
example	ir.Example	openapi:name
encoding	owning content	openapi:encoding/part/contentEncoding
oauth flows	auth scheme	openapi:flows/application
oauth flow	ir.OAuthFlow	openapi:scope
schema xml	schema	openapi:xml/attribute2
schema discriminator	schema	openapi:discriminator/mapping2
schema externalDocs	schema	openapi:externalDocs/title

Entries are keyed under the same scheme the extensions at those objects already
use — unscoped where the object has a map of its own, and scoped by the source
path from the carrier down where it rides on one. That is what lets a document
write title on its root, its tag's and its operation's externalDocs and get
three entries rather than one whose survivor depends on which lowering ran last;
the fixture writes exactly that, so the scoping is asserted rather than assumed.

The three schema sub-objects are graded as an OpenAPI object's keys (warning)
rather than as schema keywords (info), even though they hang off a schema: the
JSON Schema rule that an unrecognized keyword is legal governs the schema, and
xml, discriminator and externalDocs are OpenAPI objects the schema
vocabulary says nothing about.

One object was misclassified

components had been grouped with paths, responses and callbacks as a map
where every key is a valid entry and a census has nothing to say. It is not one:
only the map under each of its keys is the document's to name, while its own
key set is the fixed list of component kinds, which the library models as named
fields and takes a census over. Probing all four confirmed the split — the other
three fold an unrecognized key into themselves and report a type mismatch, and
their census is empty; components reports [definitions].

Two objects are deliberately left out

Both are stated at the code as well as here.

A path item (applyPathItem). The library folds a key it does not recognize
into the item's embedded operations map rather than recording it as undeclared,
so GetUnknownProperties returns nothing and there is no census to read —
confirmed at speakeasy-api/openapi v1.24.0, the version in go.mod, by reading
the census off the parsed model directly. Two things decide against the raw-node
read that would recover it:

• The key is not lost in silence, which separates it from every object above.
  Folding it produces validation-type-mismatch at error severity naming the
  key at its own pointer, and that validation is unconditional. What is lost is
  the key's value, not the fact that it was written.
• Recovering the value means reading the raw node against a path item's key
  vocabulary, and the only one this compiler owns is httpMethods — eight
  methods against the library's nine. It has no query, the method OpenAPI 3.2
  adds, and a 3.2 query operation sits in the same map as get, so the obvious
  reading would report a valid operation as an undeclared key. That vocabulary is
  what openapi: 3.2 additionalOperations are dropped entirely and silently #293 is about, so widening it here would settle that issue as a side
  effect of this one.

Filed as #377 with the reproduction and the options.

A Link Object, following the decision #345 records for its extensions. This
compiler lowers no Link Object anywhere: a response's links survive only as the
verbatim node kept on the response, and an unreferenced components/links entry
is dropped whole (verified: nothing of it appears in the compiled document). An
entry at one of the two positions would be the only trace of a construct the IR
does not model, while duplicating a value the node beside it already carries.

Test plan

• testdata/openapi/unknown_keys.yaml gains one undeclared key per object above,
  each valued with the object it was written on and each a mistake real documents
  make — a Swagger 2.0 name (definitions for schemas, a flow named
  application), a neighbouring field (title on an externalDocs,
  contentEncoding on an encoding), or a field with the number wrong (scope).
  Living under testdata/ is what runs it through the harness oracles rather
  than through assertions alone; the sweep was confirmed to reach this file by
  breaking it and watching the sweep name it.
• TestUnknownKeys_KeptAtEveryObject gains a row per object, each naming the
  Unmodeled map the entry must land on by the path the value-graph walk reaches
  it at, so an entry written to the wrong carrier fails rather than passing
  because the assertion looked only where it expected. The operation-externalDocs
  row names its carrier down to the operation, since the document writes that
  exact key too and the row is only evidence of the operation's if it cannot
  match the document's.
• TestUnknownKeys_SchemaSubObjects covers the three schema sub-objects, which
  the fixture deliberately omits: the OpenAPI dialect meta-schema closes all three
  to anything but an x- key, so an undeclared key there draws a library
  validation error, and an error diagnostic stops harness.Check before the
  oracles that fixture exists to reach. Confirmed at both 3.0 and 3.1. The keys
  are kept and announced all the same, which is what the test pins.
• Each site was proved to bite. The census was disabled at one object at a
  time and the suite re-run; all ten went red, each with the failing row naming
  that object: components, tag externalDocs, operation externalDocs, example,
  encoding, oauth flows, oauth flow, and — separately, one keyword at a time,
  since they share a reader — schema xml, discriminator and externalDocs.
• The schema-sub-object rows use assert rather than require on the lookup, so
  a keyword missing from the map reports itself and leaves the other two still
  checked. With require the first failure aborted the rest, and a row that never
  runs is no evidence about the keyword it names.
• Goldens regenerated with -update: nothing moved, including
  unwitnessed.golden.txt.
• Full gate green: gofmt, go vet, golangci-lint (0 issues), go build, and
  coverage at 100%.

Closes #357