perf: skip no-op intern retain during GC

[ianm199]

Summary

• skip retain_live_interned_strings sorting/map-retain work when the post-mark scan proves every interned short string is live
• document the evidence in the performance notes and mark the remaining intern-cache cost as record_live_interned_strings

Why this is safe

record_live_interned_strings pushes one ID for each interned short string that the marker visited. If live_ids.len() == interned_lt.len(), retaining the map would keep every entry, so sorting/deduping the ID list and walking interned_lt is a no-op. If any cached short string is dead, the lengths differ and the existing prune path still runs.

Existing tests cover both reachable and unrooted short-string cache behavior:

• interned_short_string_cache_keeps_reachable_string_until_unrooted
• interned_short_string_cache_does_not_root_unreferenced_string

Evidence

• Direct Rust A/B, best-of-20 on the same built artifacts:
  • /tmp/lua-rs-intern-baseline vs /tmp/lua-rs-intern-skip
  • gc_pressure: 0.05s -> 0.04s
  • binarytrees: 0.77s -> 0.76s
• Focused best-of-10 compare:
  • harness/bench/results/20260602T195056Z-b67d54d-compare.tsv
  • gc_pressure: 2.00x
  • binarytrees: 2.02x
  • closure_ops: 2.00x
  • table_hash_pressure: 1.33x noisy watch item
• Post-packet hotspot:
  • harness/bench/profiles/20260602T194913Z-b67d54d-gc_pressure_x300/summary.txt
  • retain_live_interned_strings drops out of the top 25 leaf frames
  • remaining intern-cache bucket: record_live_interned_strings at 2.6%
• Post-packet GC cadence:
  • harness/bench/profiles/gc-profile/20260602T195127Z-b67d54d-gc_pressure_x50/gc-rates.tsv
  • unchanged cadence: 6,336.9 collections/run
  • elapsed: 2.34061s over 50 repeated runs

Validation

• cargo test -p lua-vm interned_short_string_cache --lib
• cargo test -p lua-gc minor_collect --lib
• cargo build --release -p lua-cli --bin lua-rs
• PROFILE_REPEAT=50 bash harness/bench/gc-profile.sh gc_pressure
• PROFILE_REPEAT=300 bash harness/bench/profile-hotspots.sh gc_pressure 6
• bash harness/bench/compare.sh --runs 10 --workloads gc_pressure,binarytrees,table_hash_pressure,closure_ops
• git diff --check
• make test

Stacked on #127.