fix(proxy): strip cache_control from OpenAI format conversion

[cc10143]

Summary

Fixes two format conversion bugs in anthropic_to_openai that cause 400 errors when routing Anthropic requests to strict OpenAI-compatible endpoints (GLM, Qwen, etc.):

Bug 1: cache_control field leaks to OpenAI output

Anthropic's cache_control field (e.g. {"type": "ephemeral"}) was being passed through to OpenAI-format messages, content blocks, and tools. Strict API validators reject unknown fields with 400 Extra inputs are not permitted.

Fix: Remove cache_control passthrough from:

• System message array items
• Text content blocks within messages
• Tool definitions
• System message merge logic in normalize_openai_system_messages

Bug 2: Content array not simplified to string for single text blocks

When a message contains a single text block ([{"type": "text", "text": "Hello"}]), some strict backends expect "content": "Hello" (string) instead of "content": [{"type": "text", "text": "Hello"}] (array). The previous code kept array format when cache_control was present, which prevented simplification.

Fix: Always simplify single text block content to plain string, regardless of cache_control presence.

Related Issue

Closes #3805

Test Plan

• 57 transform tests pass (1539 total lib tests)
• New regression test test_regression_gh3805_no_cache_control_leak_to_openai validates:
  • No cache_control in any message
  • No cache_control in content parts
  • system content is plain string
  • user content is plain string
  • No cache_control in tools