chore(github-copilot): refresh model catalog

[Marvae]

Summary

• Remove Gemini 2.5 Pro and Gemini 3 Flash Preview from the GitHub Copilot catalog.
• Add Gemini 3.6 Flash and Grok 4.5 using current Copilot model metadata.

References

• Gemini 2.5 Pro and Gemini 3 Flash deprecated
• Grok 4.5 is now available in GitHub Copilot

[github-actions]

Action items

• [medium] [possible mistake] providers/github-copilot/models/gemini-3.6-flash.toml:15 - Check: Provider base_model overrides must be real host deltas, not invented capability cuts. Why: Lab google/gemini-3.6-flash inputs are text/image/video/audio/pdf, and other Copilot Gemini entries (gemini-3.5-flash, gemini-3.1-pro-preview) inherit that set and price input_audio. This file uniquely drops video/audio (and omits input_audio) with no Copilot modality citation. Action: Confirm against Copilot model metadata; keep the override only if Copilot truly lacks video/audio for this model, otherwise remove [modalities] (and restore input_audio if audio is supported).
• [medium] [possible mistake] providers/github-copilot/models/grok-4.5.toml:7 - Check: Cost values should match this host’s pricing; first-party is the baseline when Copilot mirrors lab rates. Why: First-party providers/xai/models/grok-4.5.toml and peers such as OpenRouter use cache_read = 0.3 / tier 0.6, but this entry uses 0.5 / tier 1 while keeping the same $2/$6 and $4/$12 input/output. Action: Verify Copilot’s Grok 4.5 cache pricing; if it matches xAI, set cache_read to 0.3 and tier cache_read to 0.6, and cite the source.
• [low] [possible mistake] providers/github-copilot/models/grok-4.5.toml:18 - Check: Provider limit overrides need host evidence when they diverge sharply from lab defaults. Why: Lab xai/grok-4.5 has limit.output = 500_000; this sets 128_000 (and input = 372_000) without a pricing/docs/API citation in the PR body. Action: Confirm these limits from Copilot model metadata or docs and add a citation mapped to the limit claim (or align with the verified host values).

[Marvae]

Confirmed against the current Copilot /models response:

• gemini-3.6-flash advertises only image/jpeg, image/png, image/webp, image/heic, image/heif, and application/pdf under vision. It has no audio/video capability or input_audio billing field, so the modality override and omitted input_audio are intentional.
• grok-4.5 reports cached input at $0.50 by default and $1.00 above 200K, matching GitHub's pricing table.
• grok-4.5 reports max_context_window_tokens = 500000, max_prompt_tokens = 372000, and max_output_tokens = 128000.

No catalog changes needed.