fix(models): convert input_file items that reference a file_id on the Chat Completions path

[hsusul]

Summary

The Chat Completions converter (Converter.extract_all_content) rejected any
input_file content part that did not carry inline file_data, raising
UserError("Only file_data is supported for input_file ..."). But the Chat
Completions file content part (openai.types.chat...FileFile) accepts either
inline file_data or a reference to an uploaded file_id, so an
input_file that references a file_id is fully representable and should be
converted rather than rejected.

This also fixes an internal inconsistency: the SDK's own
ToolOutputFileContent emits file_id-only input_file items (see
ItemHelpers._convert_single_tool_output_pydantic_model). A function tool that
returns ToolOutputFileContent(file_id=...) therefore produced a tool output
that the SDK's own Chat Completions converter refused to convert on the next
turn.

extract_all_content is called on tool outputs regardless of
preserve_tool_output_all_content (the non-preserve branch still calls it
before filtering to text), so this UserError aborts the run instead of letting
the caller degrade the output.

Fix (minimal): prefer inline file_data when present, otherwise fall back
to file_id. Only raise when neither is present. A file_url remains
unrepresentable on Chat Completions and still raises, and existing file_data
behavior (including the file_data-takes-precedence and optional filename
handling) is unchanged.

Before:

Converter.extract_all_content([{"type": "input_file", "file_id": "file-abc123"}])
# UserError: Only file_data is supported for input_file {...}

After:

Converter.extract_all_content([{"type": "input_file", "file_id": "file-abc123"}])
# [{"type": "file", "file": {"file_id": "file-abc123"}}]

Root cause

The input_file branch guarded on file_data only and raised for everything
else, predating (and never updated for) the file_id field that both the
Responses input_file param and the Chat Completions file part support.

Why minimal

• One converter branch changed; no new abstractions or public API changes.
• file_data path, file_data-over-file_id precedence, filename
  forwarding, and prompt-cache-breakpoint copying are all preserved.
• file_url is intentionally left unsupported (not representable on Chat
  Completions) and still raises.

Test plan

Added three focused unit tests in
tests/models/test_openai_chatcompletions_converter.py:

• test_extract_all_content_supports_input_file_file_id — converts a
  file_id-only input_file (fails on main, passes here).
• test_extract_all_content_prefers_input_file_data_over_file_id — file_data
  still wins when both are present.
• test_extract_all_content_rejects_input_file_without_data_or_id — neither
  present still raises UserError.

Validation commands (all pass):

• uv run pytest tests/models/test_openai_chatcompletions_converter.py -q → 51 passed
• uv run pytest tests/models -q → 695 passed
• uv run pytest -q → 7161 passed, 39 skipped
• uv run ruff check . → All checks passed
• uv run ruff format --check . → clean
• uv run mypy . --exclude site → Success (852 source files)
• uv run pyright → 0 errors

Compatibility: behavior only broadens (previously-raising inputs now convert);
no changes to items that already converted. No dependency or config changes.

Issue number

N/A (no OpenAI API key or live service required to reproduce).

Checks

• I've added new tests, if relevant
• I've run .agents/skills/code-change-verification/scripts/run.sh
• I've confirmed all verification steps pass
• If using Codex, I've run /review before submitting this PR