fix: recognize Bedrock inference profile ARNs for prompt caching

[giulio-leone]

Issue

Closes #1705

Problem

_cache_strategy only checked for 'claude' or 'anthropic' substrings in the model_id:

def _cache_strategy(self) -> str | None:
    model_id = self.config.get('model_id', '').lower()
    if 'claude' in model_id or 'anthropic' in model_id:
        return 'anthropic'
    return None

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 have the form:

arn:aws:bedrock:us-east-1:123456789012:application-inference-profile/abc123def456

These ARNs don't contain 'claude' or 'anthropic', so automatic caching was silently disabled with a warning:

WARNING | model_id=<arn:aws:bedrock:...> | cache_config is enabled but this model does not support automatic caching

Solution

Added detection for inference profile ARNs: when model_id starts with arn: and contains inference-profile, optimistically enable caching.

Design Rationale

• No extra API call: Avoided calling GetInferenceProfile to resolve the underlying model, which would add latency, require additional IAM permissions (bedrock:GetInferenceProfile), and create a network dependency on every converse call.
• Safe default: Currently only Anthropic Claude models support prompt caching on Bedrock. For non-caching models, the cache point is silently ignored by the Converse API — no error is raised.
• Covers both ARN types: Matches both application-inference-profile and inference-profile (cross-region) ARNs.

Testing

• Added test_cache_strategy_anthropic_for_inference_profile_arn: Tests both application and cross-region inference profile ARNs
• All 125 Bedrock tests pass (123 existing + 1 new with 2 assertions)

Changes

• src/strands/models/bedrock.py: Extended _cache_strategy to detect inference profile ARNs
• tests/strands/models/test_bedrock.py: Added test for ARN-based cache strategy detection

[giulio-leone]

Friendly ping — adds recognition of Bedrock inference profile ARNs for prompt caching, which currently only matches standard model ARNs.