fix: #2806 sanitize AnyLLM responses replay input before validation

[seratch]

This pull request fixes #2806 a multi-turn AnyLLMModel Responses-path failure where replayed history could reach any-llm local validation with adapter-only metadata and explicit None fields that ResponsesParams does not accept.

The bug shows up after prior turns have produced replay items such as reasoning blocks or tool calls that still contain fields like provider_data, placeholder response IDs, or status: None. In that state, the request fails locally during ResponsesParams(**payload) construction before it ever reaches a backend such as vLLM or Ollama.

This change narrows the fix to the adapter layer by sanitizing replayed Responses input before validation:

• strip provider_data
• drop fake response IDs
• remove recursively empty None fields
• continue filtering provider-specific reasoning items that should not be replayed

The change keeps any-llm validation in place for supported versions instead of bypassing validation globally with a shim. It also adds regression coverage for reasoning-item sanitization and for the AnyLLM Responses fetch path so subsequent tool turns continue to validate and execute with sanitized replay input.