Cut per-call overhead in the produce proxy traps

[Korijn]

Track E, PR 13 — the trap-level micro pass, the analog of observ #192/#193.

What

• Scalar fast path in _wrap: scalar reads skip the child-proxy cache lookup and the hasattr chain entirely (scalars never have a cached child proxy — cache entries only exist for container values). Inlined additionally in DictProxy/ListProxy __getitem__ for the hottest read shape.
• DictProxy.__setitem__ resolves the old value with one sentinel data.get() instead of a containment check plus a read. On plain dicts that's a wash, but when the draft wraps an observ reactive proxy (in_place=True), each of those was a trapped operation with dep tracking — two trapped calls become one.
• values()/items() iterate _data.items() once instead of re-reading every key, with the scalar check inlined; ListProxy.__iter__/__reversed__ get the same treatment plus bound locals.
• ListProxy.append and the slice read path bind hot attributes to locals and drop intermediate temporaries.

Two small tests are added for the new early-return paths (scalar list reads by index, tuple values through _wrap), keeping coverage at 100%. Behavior is unchanged; 408 tests pass, ruff and ty clean.

Benchmarks

Interleaved medians vs the memoization branch (#54), 2 runs each:

benchmark	before	after	change
produce_in_place_vs_copy_dict (writes through an observ reactive draft)	192.3µs	156.6µs	−18.6%
produce_list_many_appends[copy/in_place]	—	—	−1.8% / −2.0%
everything else	—	—	within noise (±1.5%)

An honest note on magnitude: the write-heavy plain-dict workloads were already lean (the sentinel-get trades two dict lookups for one method call, which is a wash on plain dicts), so the wins concentrate where the pass targets — read paths and proxy-backed drafts.

🤖 Generated with Claude Code

https://claude.ai/code/session_016Mu9vEBwU4fQLi8ZgkgdS2

[Korijn]

@berendkleinhaneveld I'll let you review & approve this one