Support constrained decoding

[Taka152]

This PR adds constrained decoding feature that user can provide regex or json schema to guide the decoding process. This can enable function calls or further increase the quality of the response.

The API looks like the following code, currently it only accepts regex and json_schema type, though context free grammar could be considered in the future.

# python
params.set_guidance("json_schema", json.dumps(json_schema))

// cpp
params->SetGuidance("json_schema", json_str);

The overhead brought by this feature is about 3-5% latency increase, and could further optimized after continuous decoding feature, it can even accelerate generation by jump forward tokens that is fixed in guidance.