v0.2.85

v0.2.85

- Add OPFS synchronous access handles and committed cache records, and update the bundled TVM web runtime to `0.27.0-dev0`.
- Return OpenAI-compatible `created` timestamps in seconds.
- Register service-worker handlers during startup so restarted workers can receive requests.
- Preserve recovered worker model state to avoid repeated reloads.
- Prevent hangs when initializing invalid structural-tag constraints.
- Reject NaN generation parameters.
- Render example output safely and correct cached-model `appConfig` examples.

This update continues to use the existing `v0_2_84/base` model libraries. No new model binaries are required by the validated paths.

Runtime source: [apache/tvm at 7e06fc6c1420d0188eb9d889bd74e2e1fb76e448](https://github.com/apache/tvm/commit/7e06fc6c1420d0188eb9d889bd74e2e1fb76e448).

Changes since the npm 0.2.84 source commit: [9e572d6...5f74244](https://github.com/mlc-ai/web-llm/compare/9e572d6ed95e248f29634996cd32cc8f3023d89d...5f742443179a5463e83a19f704d7c19f1f019f98).