Add constrained JSON generation for extraction and classification

[sauravpanda]

Suggested priority: P2 — application-facing capability.

Add a constrained-decoding API so applications can request structured extraction/classification results with machine-checkable output. The reviewed sampling and request APIs do not expose a JSON/schema constraint.

Scope and acceptance criteria

• Start with a documented JSON grammar and a deliberately bounded JSON Schema subset; explicitly reject unsupported schema features.
• Apply an incremental token constraint before sampling; handle tokens containing partial UTF-8 sequences and multiple grammar characters.
• Expose the feature in core generation, browser bindings/SDK, and the server request API with consistent semantics.
• Permit EOS only at valid completion points and surface unsatisfiable constraints as a clear error.
• Distinguish complete valid output from cancellation or token-budget exhaustion; never label truncated JSON as schema-valid.
• Ensure greedy, sampled, streaming, and batch paths use the same constraints; define or disable incompatible speculative paths.
• Test escaping, Unicode, nesting, enums, required fields, arrays, invalid schemas, and impossible constraints.
• Add an extraction/classification example and benchmark constraint overhead.

Initial delivery should document its supported subset instead of claiming full JSON Schema compatibility. Valid JSON/schema conformance does not guarantee factual extraction accuracy.

Implementation starting points: sampling, generation, server requests, browser bindings.