feat(server-compute): the same Rust code in the browser and on the server

New page 浏览器 vs 服务端 (algorithms). ① SHA-256, DEFLATE and sudoku run
single-threaded in the browser (the demos' own wasm) and natively on the
server, timed and checked for bit-identical results. ② The path tracer
renders on native threads and streams 16-row bands as they land; the
page compares the image with a browser render pixel for pixel.

- crates/native/compute: std-only bin running sha256/compress/sudoku/
  pathtracer natively (those crates are now cdylib + rlib); one JSON
  line per result or band. Benchmark input comes from one xorshift32
  generator, mirrored in the page's JS (first bytes pinned by a test).
- pathtracer::Tracer renders any row band byte-identical to the
  browser's progressive render (same per-(row, pass) seeds and order).
- scripts/compute_bridge.py (serve.py --mcp ... --compute): validates,
  at most 2 jobs, 60/120 s timeouts, size caps, kills a render when the
  page goes away; reuses Bridge.authorize, so cookies now use Path=/api.
- pm2 config starts with --compute.

GitNexus rated the change HIGH (serve.py dispatch, bridge auth, the
path tracer); covered by 124 Python and 252 Rust tests, a site scan
and the page e2e (76800/76800 pixels identical, 20 threads 60 ms vs
538 ms single-threaded in the browser).