refactor: extract partial eval dead code

[lbliii]

Summary

• extract const-only dead-code elimination from the partial-evaluation monolith
• share temporary inlined-body flattening across the standalone DCE pass and the main evaluator
• preserve the existing partial_evaluate() and eliminate_dead_code import contracts

Progresses #143.

Steward Notes

• consulted: compiler and testing stewards
• risk: behavior drift while moving scoping-sensitive branch elimination and temporary-node flattening
• evidence: the full partial-evaluation suite and stability gate pass; compatibility import was verified; the implementation is a phase extraction with no semantic changes
• unresolved tradeoffs: none in this slice; call/def inlining and loop unrolling remain in Split partial evaluation into smaller phase-owned modules #143

Verification

• 319 passed in the focused partial-evaluation suite
• make verify-stability (4,459 passed, 5 skipped)
• ruff check, ruff format --check, and ty check clean
• Python 3.14.2 free-threaded (sys._is_gil_enabled() == False) isolated before/after benchmarks: compile-pipeline medians flat-to-faster except 1% minimal-case variance; partial-eval medians stable, with one 125 ns dynamic microbenchmark shift attributable to timer quantization

Collateral

• no docs/changelog/API collateral: this is an internal, behavior-preserving phase extraction
• no benchmark baseline update: no reproducible regression

[Copilot]

Copilot encountered an error and was unable to review this pull request. You can try again by re-requesting a review.