fix: support Bedrock application inference profile ARNs in cache detection

[giulio-leone]

Problem

_supports_caching only checks for 'claude' or 'anthropic' in the model_id string. This works for system inference profile IDs (e.g., us.anthropic.claude-haiku-4-5-20251001-v1:0) but fails for application inference profile ARNs which contain opaque profile IDs without model names:

arn:aws:bedrock:us-east-1:123456789012:application-inference-profile/abc123

As a result, users get a warning on every API call and prompt caching is silently disabled, even when the underlying model is Claude.

Solution

• Detect application inference profile ARNs and resolve the underlying foundation model via the Bedrock control-plane GetInferenceProfile API
• Cache the resolution result to avoid repeated API calls
• Invalidate cache when model_id changes via update_config()
• Preserve original ARN casing for API calls (ARNs are case-sensitive)
• Handle API errors gracefully — if resolution fails, caching is disabled with a debug log

Tests

Added 7 new tests covering:

• ARN resolution to Claude model → caching enabled
• ARN resolution to non-Claude model → caching disabled
• API failure (ClientError) → graceful fallback
• Result caching (no repeated API calls)
• Cache invalidation on model_id change
• Original ARN casing preserved for API calls
• Empty models list response

All 128 existing tests continue to pass.

Fixes #1705

[giulio-leone]

All 5 CI checks pass. Fixes Bedrock inference profile ARN support. Ready for review.