feat: support custom Responses providers

[zidou-kiyn]

Summary

• Enable apply_patch for GPT models exposed through custom provider names when the model API is openai-responses or openai-codex-responses.
• Preserve the existing provider allowlist behavior and continue requiring model ids to start with gpt-.
• Document custom provider activation and grammar-tool fallback behavior, with tests covering both Responses API variants.

Verification

• npm run check (typecheck + biome)
• npm test (55 tests passed)
• npm pack --dry-run (release sanity)
• pi -e ./src/index.ts smoke-tested locally with a custom openai-responses provider

apply_patch impact

• Tool schema / grammar behavior is documented in README
• Workspace path safety remains covered by tests
• CHANGELOG entry added under [Unreleased]

Related issues

No existing issue directly tracks custom Responses provider activation.