fix(openai): clamp max_tokens to per-model limits to prevent overflow errors

[leseb]

What does this PR do?

Fixes BadRequestError: max_tokens is too large when clients (e.g. Claude Code) send max_tokens values that exceed what the target OpenAI model supports. For example, Claude Code requests max_tokens: 32000 but gpt-4o-mini only supports 16384.

Adds a static per-model max_output_tokens map to the OpenAI provider adapter and clamps incoming max_tokens at request time. Supports prefix matching for dated snapshot variants (e.g. gpt-4o-2024-08-06 inherits from gpt-4o). Logs a warning once per unknown model so operators know the map needs updating when new models are released.

Also populates max_output_tokens in model metadata via construct_model_from_identifier(), exposing it through the /v1/models endpoint's custom_metadata field.

Test Plan

uv run pytest tests/unit/providers/inference/test_remote_openai.py -v --tb=short

Output:

tests/unit/providers/inference/test_remote_openai.py::TestOpenAIMaxTokensClamping::test_clamps_when_request_exceeds_model_limit PASSED
tests/unit/providers/inference/test_remote_openai.py::TestOpenAIMaxTokensClamping::test_keeps_lower_request_value PASSED
tests/unit/providers/inference/test_remote_openai.py::TestOpenAIMaxTokensClamping::test_no_clamping_when_max_tokens_is_none PASSED
tests/unit/providers/inference/test_remote_openai.py::TestOpenAIMaxTokensClamping::test_does_not_mutate_original_params PASSED
tests/unit/providers/inference/test_remote_openai.py::TestOpenAIMaxTokensClamping::test_different_models_have_different_limits PASSED
tests/unit/providers/inference/test_remote_openai.py::TestOpenAIMaxTokensClamping::test_no_clamping_for_unknown_model PASSED
tests/unit/providers/inference/test_remote_openai.py::TestOpenAIMaxTokensClamping::test_dated_snapshot_model_uses_base_limit PASSED
tests/unit/providers/inference/test_remote_openai.py::TestOpenAIModelMetadata::test_construct_model_includes_max_output_tokens PASSED
tests/unit/providers/inference/test_remote_openai.py::TestOpenAIModelMetadata::test_construct_model_unknown_has_no_max_output_tokens PASSED
tests/unit/providers/inference/test_remote_openai.py::TestOpenAIModelMetadata::test_construct_model_embedding_unchanged PASSED
tests/unit/providers/inference/test_remote_openai.py::TestOpenAIMaxOutputTokensWarning::test_warns_once_for_unknown_model PASSED
tests/unit/providers/inference/test_remote_openai.py::TestOpenAIMaxOutputTokensWarning::test_all_known_models_have_limits PASSED
12 passed in 0.12s

[leseb]

With this patch i can do:

ANTHROPIC_BASE_URL="http://localhost:8321" ANTHROPIC_API_KEY="fake" claude --model "openai/gpt-4o-mini" -p "say hi in one word" 2>&1 | head
Hello.

[cdoern]

lots of new unit tests but lgtm