feat(provider): forward reasoning effort to OpenAI-compatible models

[rodboev]

Summary

SkillSpector can now forward an explicit reasoning-effort setting through its OpenAI-compatible LLM path instead of always relying on the provider or gateway default.

Closes #283

Implements the maintainer-scoped sketch in #283.

Root cause

The shared OpenAI-compatible constructor always built ChatOpenAI with a fixed kwarg set, so neither the stock OpenAI wrapper nor the nv_build wrapper had a path to forward reasoning effort. The env-var docs also had no setting for this override, which left the behavior undiscoverable even for compatible endpoints.

Diff Notes

• add optional SKILLSPECTOR_REASONING_EFFORT forwarding in the shared OpenAI-compatible constructor path
• cover configured, unset, blank, provider-wrapper, and pass-through cases in focused provider tests
• document the new env var in the existing config surfaces

Scope

This change stays in the OpenAI-compatible provider adapter path. It does not add provider-specific validation, native Anthropic support, Bedrock-specific handling, or CLI-provider changes. Unsupported provider/model combinations still fail wherever the upstream provider rejects the forwarded value.

Verification

• uv sync --all-extras passed; installed the project and 151 packages.
• Required Windows runner passed; pytest reported 7 passed, 69 deselected.
• uv run pytest tests/unit/test_providers.py -k "reasoning_effort_" passed; pytest reported 6 passed, 71 deselected.
• uv run ruff check src/ tests/ passed with All checks passed!.
• uv run ruff format --check src/ tests/ passed with 144 files already formatted.