feat(jit): specialize Set/ArrayPush on delayed-move collection rebinds

[fffonion]

Record same-local rebinds (stloc Null then Set/ArrayPush) directly in native traces, removing the call boundary that previously forced interpreter fallback.

Key changes:

• ValueInfo now carries a source_local provenance tag set exclusively at Ldloc; arithmetic, constants, and call outputs clear it, ensuring the recorder and static loop-header plan agree on which local a value came from.
• Recorder selects ArraySet / MapSet / ArrayPush specialization only when the container operand still carries the source local and that symbolic local is Null (dirty + known_type Null). Additional args that also share the moved local are rejected to prevent alias correctness issues.
• New native bridge helpers pd_vm_native_collection_set and pd_vm_native_array_push use ptr::replace to take true ownership of the container slot, perform COW/rebind/shared-impl logic, and write the result back to the same owned-temp slot, compatible with native loop-back.
• Reusable owned helpers builtin_set_owned / builtin_array_push_owned keep interpreter and JIT paths on the same semantics.
• Lowering allocates MutationArgBox temp slots for index/key/value and calls collection_set / array_push native helpers with status checking.
• Benchmark harness extended to gate per-sample native execution growth and freeze trace_attempts / recorded_traces / native traces during measurement.

Tests (all passing):

• trace_jit_specializes_same_local_array_set_through_loop_back
• trace_jit_specializes_same_local_map_set_through_loop_back (dual mutation with null-delete)
• trace_jit_specializes_same_local_array_push_through_loop_back
• trace_jit_array_set_preserves_cow_alias
• trace_jit_does_not_consume_non_moved_array_set_container

Collection rebind benchmark (release, 50K iters, width=256, 15 samples):
array: interpreter 5928 µs → JIT 1181 µs (0.20×, ~5× speedup)
map: interpreter 22822 µs → JIT 2325 µs (0.10×, ~10× speedup)

WAF benchmark sub-par because its Set calls do not use the rebind pattern; JIT trace statistics confirm the expected LoopBack shape for the micro-bench workloads.