perf(stdlib): drop per-call Vec alloc for List/String/Deque indexing 📇

[timfennis]

Summary

• The [] indexing native was uniformly registered with NativeFunc::WithVm, which forces the VM to stack.drain(args..).collect::<Vec<Value>>() on every call so it can pass &mut Vm into the callback.
• Only the Map default-function path actually needs &mut Vm (it can call user code when a key is missing). List/String/Deque indexes don't touch the VM at all.
• Split off make_get_func_simple (NativeFunc::Simple, zero-copy slice arg) for the type-specific overloads. Map and Any/Any stay on WithVm because a Map can still reach the callback path through either.
• With c12ee66 (specificity-aware overload resolution) already on master, a typed list index dispatches to the new simple overload directly.

Benchmarks

Real-world memoized DFS (AoC 2022 day 16, 5 runs, 2 warmup):

• master: 15.772s ± 0.130s
• patched: 15.328s ± 0.283s
• 1.03× faster

Profile delta (perf record, cpu_core cycles):

Symbol	master	patched
Vec::from_iter	4.46%	1.81%
run_to_depth	46.23%	48.22%
dispatch_call_with_memo	15.16%	15.17%

Total cycles drop ~4% (59.9B → 57.5B).

The []_simple and original [] show up as ~1.3% / ~1.2% each in the new profile — same total time spent in the native body, but the surrounding allocation overhead is gone for the List path.

Test plan

• cargo test — all functional tests pass, including the 27 in 007_map_and_set (covers Map default-value and default-function paths)
• cargo fmt, cargo clippy -p ndc_stdlib — no new warnings
• 2216.ndc output identical to master

🤖 Generated with Claude Code