types/client: relax think type to support model-defined thinking levels

[arrase]

Closes #743

What & Why

Models like Qwen 3.8 advertise custom thinking levels (such as 'xhigh') through Ollama's /api/show thinking capabilities metadata (introduced in ollama/ollama#18473), and the Ollama Go backend natively accepts any boolean or string for ThinkValue (types/model/thinking.go).

Previously, think was constrained to Optional[Union[bool, Literal['low', 'medium', 'high']]] across request models and client methods. Because ChatRequest and GenerateRequest are Pydantic models validated before requests are dispatched, passing valid model-defined levels like 'xhigh' (or 'max') resulted in a Pydantic ValidationError:

pydantic_core._pydantic_core.ValidationError: 2 validation errors for ChatRequest
think.bool
  Input should be a valid boolean, unable to interpret input [type=bool_parsing, input_value='xhigh', input_type=str]
think.literal['low','medium','high']
  Input should be 'low', 'medium' or 'high' [type=literal_error, input_value='xhigh', input_type=str]

This PR relaxes the think annotation to Optional[Union[bool, str]] = None across all 14 annotation sites in _types.py and _client.py, matching the Go backend and allowing any valid model-defined string thinking level while preserving full backward compatibility.

Changes

• ollama/_types.py: Updated GenerateRequest.think and ChatRequest.think from Literal['low', 'medium', 'high'] to str.
• ollama/_client.py: Updated all 12 generate and chat overloads and implementations in Client and AsyncClient to match.
• tests/test_client.py:
  • Updated test_generate_think_annotation_matches_chat to verify annotation parity across Client.chat, Client.generate, AsyncClient.chat, and AsyncClient.generate.
  • Added test_client_chat_with_think_level verifying client.chat serialization and response handling with string thinking levels (think='xhigh').
• tests/test_type_serialization.py:
  • Added test_think_model_defined_levels_serialization verifying serialization of model-defined levels ('xhigh', 'max', 'low', 'medium', 'high').
  • Added test_think_boolean_serialization verifying boolean thinking levels.

Testing

All 104 tests pass:

tests/test_client.py ...................................................
tests/test_type_serialization.py ....................
tests/test_utils.py ..........
104 passed in 1.33s

Linter and formatter check clean:

ruff check .
All checks passed!

[ParthSareen]

Thank you for this!