fix(mcp): make tool schemas compatible with Google Gemini API

[gabriel-ecegi]

Summary

The Google Gemini API rejects JSON Schema union types like ["array", "null"] in tool function declarations, causing 400 Bad Request errors when any MCP client with CK tools tries to use Gemini models.

Problem

When CK's MCP server exposes tool schemas to LLM providers, Option<Vec<String>> fields (include_patterns, exclude_patterns) generate:

{
  "type": ["array", "null"],
  "items": { "type": "string" }
}

This is valid JSON Schema but Gemini's API validates strictly:

GenerateContentRequest.tools[0].function_declarations[N].parameters.properties[include_patterns].items: 
  field predicate failed: $type == Type.ARRAY

Fix

Add #[schemars(with = "Vec<String>")] to all 8 Option<Vec<String>> fields across the 4 search request structs (SemanticSearchRequest, RegexSearchRequest, HybridSearchRequest, LexicalSearchRequest).

This generates "type": "array" instead of "type": ["array", "null"] while keeping the Rust type Option<Vec<String>> unchanged — fully backward compatible.

Tested

• ✅ google/gemini-3.1-pro-preview — responds correctly
• ✅ google/antigravity-gemini-3-flash — responds correctly
• ✅ Anthropic models — still work (they were already fine)
• ✅ cargo test passes
• ✅ cargo build --release succeeds