Replace SUPPORT_REASONING_EFFORT_MODELS with litellm's model metadata plus the config override

Summary

pr_agent/algo/__init__.py:465 keeps a hand-maintained list of 14 model ids that may receive reasoning_effort, read only by litellm_ai_handler.py:2173. Every new reasoning model needs a PR to extend it (#3413, #3309, #3459 are the last three), and a model missing from it has its configured effort dropped silently.

Proposal

Decide from litellm's own metadata (litellm.supports_reasoning(model), in the bundled cost map pinned by #3475) and keep config.additional_reasoning_effort_models from #3460 as the operator escape hatch for endpoints litellm does not know. Delete the list once both paths cover today's entries.

Cases the replacement must keep passing

• The six bare Grok ids resolve only with the xai/ prefix in litellm (measured: supports_reasoning is true for 8 of the 14 entries, false for all six Grok ids as listed), so the lookup needs the provider-prefixed form, including openrouter/xai/... and :nitro-style suffixes.
• Gemini 2.5 keeps its thinking-budget mapping; Gemini 3.x follows #2998 and #3434.
• gpt-5.x clamping from #3273 is unaffected.
• test_litellm_reasoning_effort.py is the regression suite; the 26-model matrix from the #3460 review is the acceptance check (identical kwargs before and after).

Depends on #3460 landing first.