fix(litellm): record every tool_result block of an anthropic message, not just the last

[Harsh23Kashyap]

Fixes #3817

When a user message carries several tool_result blocks (the standard reply to parallel tool calls), litellm's anthropic messages extractor writes every block to the same two keys, message.content and message.tool_call_id, at the message's index. Only the last block reaches the span; the earlier tool results are dropped from the trace.

This change records each tool_result block as its own input message with message.role = "tool", message.content, and message.tool_call_id:

• A message holding only tool_result blocks expands into one tool message per block instead of a single "user" message that keeps only the last result.
• Mixed contents are left as they are, as are string contents, the system prompt path, and the output message path.

Same fix shape as #3810 (anthropic), #3812 (vertexai), and #3814 (google-genai), per spec/tool_calling.md.

Test plan:

• New regression test test_create_input_multiple_tool_result_blocks in tests/test_anthropic_messages.py; fails on main and passes with this change.
• pytest tests/test_anthropic_messages.py - 24 passed, including the existing content-block tests, which are unaffected.

[ialisaleh]

I reviewed the PR, and the change fixes the reported case by recording each parallel string tool result as a separate tool message with its content and tool_call_id. The regression test covers both results. As a follow-up, we can check list-valued tool_result.content, since the new path changes how it is recorded and may omit some nested block types. Thanks!