fix(deepseek): map minimal reasoning effort to low - #6132

DeepSeek's Chat Completions requests currently send reasoning_effort="minimal" together with thinking.type="disabled", even though DeepSeek documents minimal as an alias for low-effort thinking.

Normalize both the semantic and wire effort to low in the DeepSeek request builder. This enables thinking consistently and lets the existing history handling backfill missing reasoning_content for tool calls. Explicit none still disables thinking, and an unspecified effort keeps the provider default.

The production change is three lines. Tests cover minimal, low, none, and the default, plus existing reasoning history and missing tool-call reasoning. The provider registry and other providers' effort mappings are unchanged.

Fixes #6122.

Validation on Windows / Python 3.13:

• Before the production fix, the updated regression selection failed in three cases: the conflicting effort and missing history reasoning for plain/prefixed deepseek-flash.
• python -m pytest tests/providers/test_openai_compat_spec.py tests/providers/test_mistral_provider.py tests/providers/test_xiaomi_mimo_thinking.py -q: 194 passed.
• python -m pytest tests/providers/test_model_api_presets.py -q -k deepseek: 20 passed.
• ruff check nanobot tests conftest.py: passed.
• Scoped BasedPyright check of nanobot/providers/openai_compat_provider.py: 0 errors, 0 warnings.
• Python compile checks and git diff --check: passed.

Verification used offline fixtures and isolated local dependencies. No live provider calls were made; the full Python suite and CI dependency/platform matrix were not run locally.