feat: cross-validation harness and performance regression tracking

[tbitcs]

Cross-validation harness and performance regression tracking

Cross-validation test suite ( ests/python/test_cross_validation.py)

• 12 golden test vectors in ests/vectors/ covering all major evaluation paths:
  basic eval, safety guard ordering, expression opcodes (scale, accumulate, clamp with saturation),
  div-by-zero, staleness, mode transitions, fault raise/clear, delta operators, condition groups,
  and INT32 saturation
• Determinism test: runs 3 vectors 100 times each, asserts identical output
• Compile-to-C test: compiles each vector model to C source and validates required symbols
  (ARBITER_generated_model, ARBITER_MODEL_HASH)

Performance regression tracking

• ools/parse_benchmark.py: parses Twister benchmark logs, extracts ns/tick timing for
  hand-coded and engine variants from PID and Kalman benchmarks, prints summary table
• CI workflow updated with "Benchmark timing summary" step after Twister benchmarks
  (log-only for now — no fail threshold until baseline data is collected)

Test results

• All 180 tests pass (python -m pytest tests/python/ -v)
• 27 new cross-validation tests (12 golden vectors + 3 determinism + 12 compile-to-C)

Conversation: https://app.warp.dev/conversation/069a3382-5169-4a07-8e15-ea4862186401
Run: https://oz.warp.dev/runs/019e8993-e9a4-7058-9fea-abfbf925ac06

This PR was generated with Oz.