fix(gc): avoid closure side-table GC deadlock

[andrewtdiz]

Summary

Fixes the gc_write_barrier_stress::tenured_mutation_stress hang that blocked PR #4928's GitHub Actions cargo-test run after cancellation, separately from the optimized-lib freshness change.

Root cause: during minor copying GC, scan_closure_dynamic_props_roots_mut held the closure dynamic-property/prototype side-table mutexes while visiting JS value slots. Visiting those slots can evacuate a closure, and the move hook closure_dynamic_props_owner_moved re-enters the same side-table lock, deadlocking the runtime inside the first explicit gc() after tenured objects point at nursery values.

The fix detaches each closure side-table entry before invoking GC visitor callbacks, then merges it back under the forwarded owner after callbacks complete. It also prevents non-closure receivers from entering closure-only Function.prototype fallback, which the stress exposed after the deadlock was removed.

Verification

• Reproduced the original hang locally with a process-group timeout wrapper: cargo test -p perry --test gc_write_barrier_stress tenured_mutation_stress -- --nocapture timed out after 120s and left no main_bin/cargo orphan after wrapper cleanup.
• Sampled a no-auto optimized reducer at the post-mutation gc() stall. Stack showed scan_closure_dynamic_props_roots_mut -> visit_nanbox_f64_slot -> move_young -> closure_dynamic_props_owner_moved -> CLOSURE_PROPS.lock.
• cargo test -p perry-runtime closure::dynamic_props::tests_1802 -- --nocapture => 4 passed.
• cargo build -p perry-runtime => passed, existing warnings only.
• No-auto fixture after fix: RUN_STATUS 0 timed_out False elapsed 26.0, stdout BARRIER_STRESS_OK.
• Normal-mode focused test: cargo test -p perry --test gc_write_barrier_stress tenured_mutation_stress -- --nocapture => 1 passed, finished in 150.24s.
• Full stress file: cargo test -p perry --test gc_write_barrier_stress -- --nocapture => 2 passed, finished in 10.77s.
• cargo fmt --check --package perry-runtime --package perry => passed.
• git diff --check => passed.
• Process audit found no matching leftover gc_write_barrier_stress, main_bin, perry compile, or perry-auto process from this worktree.

CI note

PR #4928's cancelled cargo-test appears blocked by this separate GC/runtime stress issue, not by the optimized-lib freshness assertions. After this lands, rerun PR #4928's cancelled check.