SH3-5 round 3: Rust AST perf — Copy ShapeCk + step-level FIRST pre-filters: 7.4x → 5.9x vs CST

- ShapeCk is now Copy: drop lists_len/holes_len + Rc suppress snapshots;
  suppress writes go through set_/take_ setters that append old values to
  suppress_log, restore replays back to a log watermark — no Rc refcount
  churn per checkpoint (~4.4M ck/restore pairs on the 2MB bench) (O4)
- Step-level FIRST pre-filters (O5): per-rule leading FIRST built once
  per emit (buildRustShapeRuleFirst; null = unknown); sound seq FIRST via
  rustShapeFirstOfSteps (fully-nullable seqs get no guard). alt branch
  pick, star exit walk, opt, and sep leading element skip the doomed
  ck + walk + restore when the current token misses FIRST
- emitRustAstRdAltSteps: drop unused listsVar/holesVar plumbing
- Gates: test/shape-rust.ts 32/32; accept-corpus TS↔Rust iso 0/500,
  toy 0/2254, typescript accept 2000 cases shapeDiv=0
- 2MB paired: median ast/cst 5.86x (was 7.43x, −21%)

Author: johnsoncodehk