Add benchmark CI for performance regression detection

[m96-chan]

Summary

A bench.rs file exists in nxpu-e2e-tests but there is no CI integration to track performance over time or detect regressions.

Motivation

• Optimization passes and backend changes can silently degrade compile-time performance
• Without baseline tracking, regressions go unnoticed until users report slowness
• Automated benchmark comparison on PRs catches issues early

Proposed Approach

1. Benchmark framework: Migrate to criterion.rs for statistical benchmarking
2. CI integration options:
   • bencher.dev — Free for open-source, GitHub integration
   • github-action-benchmark — Stores results in gh-pages branch
   • Custom: store baseline in artifacts, compare with critcmp
3. Benchmark targets:
   • Parse time (each example WGSL file)
   • Optimization pass time (per pass and total pipeline)
   • Backend emit time (per backend)
   • End-to-end compile time (WGSL → output format)

Acceptance Criteria

• criterion.rs benchmarks for key pipeline stages
• CI job that runs benchmarks and compares against baseline
• PR comments showing benchmark diff (or link to dashboard)
• Alert/failure on significant regressions (e.g., >10% slowdown)