fix: #1704 Preserve thinking blocks in Anthropic conversations with tool calls

[habema]

Fixes #1704

Problem:
The SDK was dropping reasoning items when converting conversation history to API messages, causing Anthropic to reject requests with:

messages.1.content.0.type: Expected `thinking` or `redacted_thinking`, but found `tool_use`

Root cause:
Converter.items_to_messages() was skipping reasoning items entirely, violating Anthropic's requirement that assistant messages with tool calls must start with thinking blocks.

Technical changes:

1. Enhanced InternalChatCompletionMessage to carry thinking_blocks alongside reasoning_content

2. Updated LitellmConverter.convert_message() to extract full thinking blocks from LiteLLM responses:

   • Handles both dict and object formats of thinking blocks
   • Preserves the cryptographic signature field from Anthropic
   • Safely converts to dict format for storage

3. Modified Converter.message_to_output_items() to store complete thinking block structure:

   • Stores thinking blocks in ResponseReasoningItem.content as Content objects
   • Maintains backward compatibility with existing reasoning content

Repro script available in original issue.

[habema]

make old_versions_tests is not failing locally