Fix MCP call_tool ignoring per-server tool_timeout

[PaoloC68]

Summary

• call_tool in MCPClientBase always used the global mcp_client_tool_timeout setting, ignoring the per-server tool_timeout override
• This caused tools with long execution times (e.g., Perplexity deep research at 5-20 min) to timeout prematurely at the global default (120s), even when the server config specified a higher timeout
• Applied the same resolution pattern already used for SSE/HTTP transport connections (line ~1110): prefer per-server tool_timeout, fall back to global setting

The Bug

# BEFORE (line 984) - always uses global timeout
read_timeout_seconds=timedelta(seconds=set["mcp_client_tool_timeout"])

# AFTER - per-server override takes precedence
tool_timeout = self.server.tool_timeout or set["mcp_client_tool_timeout"]
read_timeout_seconds=timedelta(seconds=tool_timeout)

The per-server tool_timeout was already being used for the transport-level SSE read timeout, but not for the actual MCP protocol call_tool operation. This one-line fix makes them consistent.

Impact

Any MCP server with a custom tool_timeout in its config was silently having that value ignored during tool execution. Only the transport connection respected it.

Testing

• Verified the fix resolves Perplexity deep research timeouts (configured tool_timeout: 1200 for 20-minute research queries)
• No changes to behavior for servers without a custom tool_timeout (falls back to global as before)