feat: add wasm web support

[jhen0409]

Summary

Adds browser WebAssembly support for llama.node through a new @fugood/node-llama-wasm package. The package keeps the high-level llama.node API shape for browser use, loads GGUF models by URL by default, supports session save/load with buffers and URLs, and includes CPU plus WebGPU WASM artifacts.

The CPU path now auto-selects the pthread artifact on cross-origin isolated pages and defaults n_threads to min(4, navigator.hardwareConcurrency). The WASM runtime runs in a dedicated worker by default so model loading, tokenize/detokenize, completion, state, embedding/rerank, multimodal staging, and bench calls do not block the UI thread.

Validation

• node --check packages/node-llama-wasm/index.js
• node --check packages/node-llama-wasm/worker.js
• node --check scripts/build-wasm-package.js
• node --check scripts/build-wasm-docker.js
• npm run typecheck
• npm pack --dry-run ./packages/node-llama-wasm
• Browser HTML harness with Gemma 3 270M CPU threaded: load/tokenize/detokenize/state/completion/tool formatting passed
• Browser HTML harness with Gemma 3 270M WebGPU: load/tokenize/detokenize/state/completion/tool formatting passed
• Apple Silicon Docker smoke: npm run build-wasm-docker selected emscripten/emsdk:4.0.14-arm64 (linux/arm64) and copied the CPU WASM artifact
• Browser smoke with the arm64-built CPU artifact: tiny model load, tokenize/detokenize, worker runtime, and bench passed

Mac Docker Benchmark

Clean CPU WASM build on Apple Silicon Docker Desktop with JOBS=4 and isolated Emscripten caches:

• emscripten/emsdk:4.0.13 with --platform linux/amd64: 1268.44s
• emscripten/emsdk:4.0.14-arm64 with --platform linux/arm64: 122.51s

The native arm64 image is about 10.4x faster for this build on Mac, so npm run build-wasm-docker now selects 4.0.14-arm64 on arm64 hosts.

Workflow Dispatches

Dispatched .github/workflows/build-release.yml on branch codex/add-wasm-web-support with upload-artifacts=YES and publish=NO:

• target=wasm: run 26815857397, build-wasm job - success
• target=darwin-arm64: run 26815867290, node-llama-darwin-arm64 job - success