Use model's generation_config.json for default sampling parameters

[eyupcanakman]

Fixes #140.

Models like Phi-4 ship a generation_config.json with sampling defaults (temperature, top_p, etc.) but mlx-lm only read eos_token_id from it. Now generate, chat, and server read these values and use them when the user does not specify an explicit override.

The priority chain is: user CLI arg > generation_config.json > hardcoded default.

The server resolves defaults lazily per request so model hot-swapping picks up the new model's config correctly.

Also adds --min-p and --top-k CLI args to chat.py (already present in generate.py) so generation_config values for those keys are not silently ignored.

[Thump604]

The three-tier priority (per-request > generation_config.json > hardcoded defaults) and the lazy resolve_default() on ModelProvider are well designed. One edge case: some HF configs set do_sample: true with temperature: 1.0, which effectively means "use default sampling." The current do_sample: false -> temp: 0.0 mapping is correct, but when do_sample: true the code probably shouldn't inject the config's temperature since 1.0 is just the HF default, not an intentional override.