fix(models): stop retrying permanent context overflow errors

[daleselaji-dev]

Summary

Fixes #921.

Context-window overflow responses are permanent request errors, but the model retry loop currently treats them like transient provider failures. A single oversized prompt can therefore resend the full history up to ten times before surfacing the original error.

This change:

• classifies the documented context_length_exceeded and context_window_exceeded 400 responses as non-retryable;
• applies the same behavior to OpenRouter's chat, text, and Responses API models and to Requesty;
• adds a focused regression test that verifies each model makes exactly one request.

Root cause

Provider-specific model classes only excluded authentication errors and KeyboardInterrupt from Tenacity retries. Context overflow was raised as the generic API exception, so the retry policy resent an unchanged request even though retrying cannot reduce its size.

Validation

• python -m pytest -q tests/models/test_context_length_retry.py tests/models/test_openrouter_textbased_model.py — 11 passed
• real local HTTP 400 server reproduction — exactly 1 request
• Ruff format and lint checks — passed
• python -m compileall -q src/minisweagent/models tests/models/test_context_length_retry.py — passed

The broader model test collection still has six unrelated failures in tests that import the optional litellm dependency; no changed code is involved in those failures.

Compatibility

Only recognized context-overflow 400 responses change behavior. Authentication, rate-limit, network, and other provider errors keep their existing retry policies.