openai: support custom apply_patch Responses tools

[pd95]

Summary and motivation

Codex now exposes apply_patch as a Responses API custom/freeform tool. That
tool carries a patch as raw text, while Ollama's internal chat tool interface
uses JSON function arguments. Without a compatibility bridge, Ollama cannot
pass the declaration, model call, tool output, and continued conversation
through /v1/responses in the shapes expected by coding agents.

What this changes

This PR adds a deliberately narrow bridge for apply_patch:

• converts a directly declared top-level custom tool into an internal function
  with one required string field, input;
• exposes the expected free-form Lark grammar on the Responses wire format;
• converts a valid generated patch into custom_tool_call events in streaming
  and non-streaming responses;
• accepts custom_tool_call and custom_tool_call_output history so a client
  can execute the patch and continue the turn;
• retains call/output pairs together during compaction, failing closed rather
  than preserving an orphaned half;
• rejects nested declarations, dynamically discovered custom tools, and custom
  tool names other than apply_patch;
• leaves malformed patch output visible through the ordinary fallback instead
  of silently discarding or misclassifying it.

The server translates protocol data only. It does not apply patches or execute
model-produced commands.

Dependency and scope

This PR depends on Responses namespace PR #8 and changes only Responses
request/output conversion, compaction, and focused middleware coverage.

Compare:
feature/responses-namespace-tools-upstream...feature/responses-custom-apply-patch-tool

Validation

Conversion, schema, grammar, collision, history, ordering, compaction,
malformed-output, streaming, and mixed-tool middleware tests pass. The required
Linux test/lint/build and integrated Darwin gates also pass.

Refresh and AI assistance

Adapted onto Ollama v0.34.0; head
4cec20fde16a9aac9ff800c7a77463e38b0fd2ed. The final commit repairs a test
fixture so it declares the custom tool it exercises. Codex materially assisted
implementation review, tests, refresh validation, and this PR description.