[SCH-7477] cancel in-flight prewarm before releasing leases on shutdown

[schematic-orbit]

shutdown() stopped the sweeper and released the per-process leases, but left the fire-and-forget prewarm work running. A prewarm acquire that landed after release_all_local_leases() listed the store installed a lease nothing was left to release, so its credits stayed held against the company balance until the server expired it — five minutes by default.

The load-bearing part is tracking the single-flight acquire. _single_flight registers the task and then await asyncio.shield(task), so cancelling a prewarm cancels the shield, not the acquire underneath, and the finally drops the registry entry as the caller unwinds. The acquire was therefore tracked nowhere at all and went on to call _lease_store.replace() behind the release. It now joins the manager's drain set alongside the registry entry: the registry dedupes concurrent callers, the drain set is what a close waits on, and they have different lifetimes.

shutdown() now goes stop → cancel → drain → release. It cancels the prewarms in _background_tasks, awaits LeaseManager.drain() (bounded by a new internal SHUTDOWN_DRAIN_TIMEOUT of 5s, since a shutdown that hangs is worse than a hold the server expires in DEFAULT_LEASE_DURATION), and only then releases. Cancel and drain run for a shared backend too — that work must not outlive the client — while the release stays gated on not self._lease_backend_shared. _spawn is a no-op past stop(), so extend_in_background starts nothing new, and _prewarm_one returns early while the client is shutting down, which covers a caller awaiting client.prewarm(...) directly rather than through identify(prewarm=...).

Only the in-memory store releases on shutdown, so this is single-process deployments; a Redis-backed store still releases nothing.

Before, with an acquire on the wire when shutdown begins:

>       assert client._lease_store.list_leases() == []
E       AssertionError: assert [LeaseState(lease_id='lse_1', company_id='co_1', ...)] == []
E         Left contains one more item

After, the lease is drained, found, and released — release_credit_lease is awaited for lse_1 and the store is empty.

Notes, not in the diff:

• The Node half of the ticket (close() in schematic-node) is not here — one repo per change. It needs a closing flag resolveCompanyIdWithWait's poll loop checks, a tracked set for the voided this.prewarm(...) in identify, and a drain() on CreditLeaseManager awaiting inflightAcquire/inflightExtend plus the voided release and background extend. Promises are not cancellable, so Node drains where Python cancels.
• A check() racing shutdown() can still install a lease after the release. Gating acquire_if_needed on _stopped would change what a mid-shutdown check returns (fail-closed denies), which is a call for a human and not what this ticket claims.
• No conformance vector: conformance/SPEC.md's manager-level ops have no prewarm and no client-close op, so covering this there means a spec change shared with schematic-node.

https://linear.app/schematic/issue/SCH-7477/sdk-shutdown-does-not-cancel-in-flight-prewarm-before-releasing-leases