fix: allow ThinkingBlock/RedactedThinkingBlock in RequestContentBlock

[Ch33zig]

Fixes #71.

RequestContentBlock (used to validate messages[].content for both user and assistant roles) only allowed TextBlock | ImageBlock | ToolUseBlock | ToolResultBlock. ResponseContentBlock already includes ThinkingBlock and RedactedThinkingBlock for outgoing responses, but the request-side union never got the same treatment.

Repro: proxying client.messages.create(model="claude-sonnet-5", thinking={"type": "adaptive"}, ...) through /claude in a multi-turn tool-use loop. The model returns a thinking block on turn 1; once that block is echoed back in messages[] on turn 2 (required for tool-use continuation per the Messages API), the proxy rejects the request with a 422:

{"type": "error", "error": {"message": "body -> messages -> 1 -> content -> str: Input should be a valid string; body -> messages -> 1 -> content -> list[tagged-union[TextBlock,ImageBlock,ToolUseBlock,ToolResultBlock]] -> 0: Input tag 'thinking' found using 'type' does not match any of the expected tags: 'text', 'image', 'tool_use', 'tool_result'"}}

This isn't model-specific — any config that returns thinking blocks (adaptive or budget_tokens-style extended thinking) and gets used in a tool-calling loop hits this, since the Messages API requires thinking blocks to be replayed unchanged in conversation history.

Fix: add ThinkingBlock | RedactedThinkingBlock to RequestContentBlock, mirroring ResponseContentBlock.