Add anthropic_cache_messages and CachePoint stripping in AnthropicModel

[Wh1isper]

closes #3441
closes #3467

[Wh1isper]

One more thing, if enable anthropic_cache_all, we should ignore all CachePoint in message history

[DouweM]

One more thing, if enable anthropic_cache_all, we should ignore all CachePoint in message history

@Wh1isper Ignore all, or only respect the final 4 i.e. #3467?

[Wh1isper]

@DouweM You're right, we can handle this more generically:

1. When enable anthropic_cache_all, mark the last message as a cache point
2. Retain the last four cache points(2-4 for message, 1 for instruction, 1 for tool definition)