🧩 Expand modelOverrides for mode, tokens, and params (#124)

## Summary

Extends opt-in `modelOverrides` so self-hosted LiteLLM gateways with
incomplete `/model/info` cards can be corrected in settings (no local
metadata proxy).

Closes #123

## Changes

- **Mode / token fields:** override `mode`, `max_input_tokens`,
`max_output_tokens`, `max_tokens`, `context_window_tokens` at discovery
time
- **`supportedOpenaiParams`:** full-list replace onto
`supported_openai_params` (was parsed but unused)
- **Param filtering:** explicit `supported_openai_params` wins over
static family denylists
- Settings schema + README examples
- `modelOverrides.json` stays empty (no bundled cards)

## Notes

- Requires `enableModelOverrides: true`
- `forceResponsesEndpoint` still forces `chat` → `responses` after
overrides

## Validation

- `npm run compile` pass
- `modelOverrides` + parameter filtering tests pass
- `npm run lint` / `format` pass

---------

Co-authored-by: amwdrizz <amwdrizz@users.noreply.github.com>