fix(provider): correctly parse GitHub Copilot models response

[amitksingh1490]

Problem

Users report that the GitHub Copilot provider lists models that fail with model_not_supported when selected (e.g. on the GitHub Student / free plan), and that model capability metadata (context length, tools, reasoning, vision) is missing for all Copilot models.

Root cause: the Copilot /models endpoint does not use the standard OpenAI models schema. It nests metadata under capabilities, limits, and policy objects. Forge parsed it with the generic OpenAI ListModelResponse, which:

1. Listed policy-disabled models — models with policy.state == "disabled" for the account's subscription appear in :models but return 400 model_not_supported when used.
2. Listed non-chat models — embedding models like text-embedding-3-small cannot chat at all.
3. Dropped all capability metadata — context length, tool support, reasoning, and vision showed as unknown.

Note: the "Auto" option seen in VS Code is a client-side VS Code feature, not an API model. The Copilot API has no auto model id (verified: requesting "model": "auto" returns 400 model_not_supported). Specific model ids continue to work directly via the API.

Solution

Add a Copilot-specific DTO (CopilotListModelResponse / CopilotModel) that mirrors the real schema, and use it in OpenAIProvider::inner_models() when the provider is github_copilot:

• Deserializes the Copilot models schema (capabilities, limits, policy, supported_endpoints)
• Filters out policy-disabled models, non-chat models, and entries missing token-limit or tool-call metadata (mirrors the checks official clients perform)
• Maps context window, tool calls, parallel tool calls, reasoning (adaptive thinking / reasoning effort / thinking budget), and vision to the domain model

The filtering rules match what opencode applies in its github-copilot plugin.

Verification

• Validated against a live Copilot account (student/free plan): the filter yields 36 usable models; previously listed but broken entries (e.g. gpt-5.5, claude-opus-4.8, embeddings) are removed
• Confirmed chat completions succeed (200) for filtered-in models (gpt-5-mini, claude-sonnet-4.6, gpt-4o, gpt-4.1) and fail (400) for filtered-out ones (gpt-5.5)
• 7 new unit tests built from real API responses
• cargo check, cargo clippy, and all forge_app + forge_repo tests pass (999 tests)

Co-Authored-By: ForgeCode noreply@forgecode.dev