feat: implement multithreaded WASM support, WebGPU fallback, and asse… - #108

Summary

The web demo took up to 10 minutes to become usable because ~400 MB of ONNX
weights were being downloaded inefficiently. This PR rewires the dev server,
loader, and ONNX Runtime config so the first load is several times faster and
subsequent loads are near-instant.

What was wrong

• Sequential downloads. The 4 ONNX models (~400 MB total) were fetched one
  after another in a for loop.
• Double-download on WebGPU fallback. If WebGPU init failed, the code
  called the loader a second time with executionProviders: ['wasm'],
  re-downloading every model — ~800 MB worst case.
• No browser caching. The dev-server middleware piped files with no
  Content-Type, Content-Length, ETag, or Cache-Control, so every
  refresh re-fetched the full 400 MB and the UI could not show real progress.
• WASM threads disabled. No COOP/COEP headers meant SharedArrayBuffer
  was unavailable, so onnxruntime-web ran single-threaded.

Changes

web/helper.js

• loadOnnx now fetches the model itself as an ArrayBuffer (with streaming
  byte-level progress and cache: 'force-cache') and passes the buffer to
  InferenceSession.create. ORT can switch execution providers without
  re-downloading the weights.
• loadTextToSpeech downloads all 4 models in parallel via Promise.all
  instead of sequentially.
• Progress callback now reports aggregated MB downloaded / total MB.

web/main.js

• Detect WebGPU once with navigator.gpu.requestAdapter(), then call the
  loader once with executionProviders: ['webgpu', 'wasm']. ORT falls
  back per-op internally — no second download.
• Configure ort.env.wasm up front: SIMD on, numThreads set from
  navigator.hardwareConcurrency (capped at 4) when the page is
  cross-origin isolated.

web/vite.config.js

• Static-asset middleware now sets Content-Type, Content-Length, ETag,
  Accept-Ranges, and Cache-Control: public, max-age=31536000, immutable,
  with 304 Not Modified short-circuit and HTTP Range support.
• New COOP/COEP plugin sets Cross-Origin-Opener-Policy: same-origin and
  Cross-Origin-Embedder-Policy: require-corp on all dev/preview responses,
  enabling SharedArrayBuffer (required for multi-threaded WASM).

Test plan

• npm run dev — first load: 4 .onnx requests fire in parallel in the
  Network tab; status shows live MB progress.
• Refresh the page — all 4 models return 304 Not Modified (or come
  from the disk cache); UI is ready almost immediately.
• On a machine with WebGPU: backend badge reads WebGPU, no second
  download in the Network tab.
• On a machine without WebGPU: falls back to WASM with a single download.
• In the console: self.crossOriginIsolated === true and
  ort.env.wasm.numThreads > 1.
• Generate speech end-to-end and verify the audio plays.