fix(models): stop retrying context window errors

[swayamg20]

Summary

• Classify only HTTP 400 context-window overflows using canonical provider error fields and a conservative message fallback.
• Raise provider-specific API-error subclasses across the OpenRouter tool-call, text-based, and Responses adapters, plus Requesty, so permanent failures abort immediately.
• Preserve existing retries for unrelated 400 responses, transient 5xx responses, and network errors.

Closes #921

Testing

• Real local HTTP-server tests verify context overflows make one request while other API errors retain the configured retry count.
• pytest -q tests/models/test_context_window_errors.py tests/models/test_openrouter_textbased_model.py (16 passed)
• pytest -q -n auto (552 passed, 76 skipped)
• Ruff check and format validation on all changed files
• All applicable Python pre-commit hooks on all changed files

[EvolveAegis]

Independent corroboration from a black-box PoC (no source patching): against the shipped default stop_after_attempt(10), a context_length_exceeded HTTP 400 from the provider goes through up to 10 total attempts (1 initial plus 9 retries) before giving up, while a 401 correctly aborts at attempt 1. We also verified the bug is provider-symmetric: OpenRouter and Requesty share the same retry.py, so both exhibit the same retry waste. This PR's is_context_window_error() plus a dedicated ContextWindowError in abort_exceptions moves the 400 from "retry to budget" to "abort at attempt 1", and the added test proving generic 400/500/token_limit_exceeded still retry confirms it doesn't over-suppress. One scope note: only context_length_exceeded-shaped 400s abort; other permanent 4xx still retry to the budget, worth a follow-up.

[swayamg20]

Thanks for independently validating this. Agreed that broader permanent-4xx handling should remain a follow-up, this PR intentionally scopes the non-retry behavior to recognized context-window errors.