Split partial evaluation into smaller phase-owned modules

[lbliii]

Context

partial_eval.py is one of the highest-risk files in the project: it owns compile-time evaluation, constant folding, dead-code elimination, loop unrolling, call inlining, and flattening. It is large enough that unrelated changes are likely to collide and tests are similarly concentrated.

Evidence

• src/kida/compiler/partial_eval.py is 2,307 lines.
• PartialEvaluator spans about 1,671 lines.
• _transform_expr alone is about 298 lines.
• tests/test_partial_eval.py is 3,527 lines.

Proposed Scope

• Split the implementation along behavior boundaries, such as expression evaluation, control-flow transforms, call/def inlining, loop unrolling, and flattening.
• Preserve the public partial_evaluate_template() entrypoint.
• Split or regroup tests around the same behavior boundaries.
• Include benchmark/no-benchmark rationale because this is compiler hot-path-adjacent.

Done When

• Partial-eval changes can be reviewed in smaller files with clear ownership.
• Existing partial-eval tests still pass.
• Compiler benchmark expectations are unchanged or documented.