fix(models): handle stalled litellm tool-call streams

[snvtac]

Fixes #874.

Summary

• add a controlled ProviderTimeout exit for LiteLLM provider read timeouts
• consume LiteLLM chat stream wrappers into a normal response so streamed tool-call deltas are parsed by the existing tool-call path
• apply the default read timeout only for stream=True calls, preserving explicit timeout / request_timeout overrides and leaving non-stream requests on LiteLLM defaults
• cover stalled and successful streamed tool-call responses, plus clean agent exit behavior

Tests

• MSWEA_SILENT_STARTUP=1 .venv/bin/python -m pytest -q tests/models tests/agents
• PRE_COMMIT_HOME=.cache/pre-commit .venv/bin/pre-commit run --files src/minisweagent/exceptions.py src/minisweagent/models/litellm_model.py src/minisweagent/models/litellm_response_model.py src/minisweagent/models/litellm_textbased_model.py tests/agents/test_default.py tests/models/test_litellm_model.py tests/models/test_litellm_textbased_model.py tests/models/test_litellm_response_model.py