Add mcp_server_name to tool metadata in convert_mcp_tool_to_langchain_tool

[RobinBaldeo]

Closes #484.

When MultiServerMCPClient.get_tools() aggregates tools from multiple MCP servers, the server identity is currently only recoverable by parsing the prefix string from tool.name (and only when tool_name_prefix=True). If tool_name_prefix=False, the server identity is lost entirely after the aggregate list is returned.

This PR writes tool.metadata["mcp_server_name"] at the point where each tool is constructed in convert_mcp_tool_to_langchain_tool. The server_name is already threaded end-to-end through the call chain (get_tools → load_mcp_tools → convert_mcp_tool_to_langchain_tool), so this is an additive change with no signature impact.

This change is purely additive — it adds a new top-level key (mcp_server_name) to tool.metadata. Existing keys and consumers reading them are unaffected.

Use case: building a framework where MCP tools from multiple servers need attribution for telemetry and UI rendering. Parsing tool.name.split("__")[0] is fragile and doesn't work when tool_name_prefix=False.

Collision behavior: if an MCP server's annotations field contains a key named mcp_server_name, it will be overwritten by the client-configured server key (the key in self.connections). This is intentional — the metadata tag reflects client-side configuration, not a server self-declaration, so the value is consistent regardless of server behavior.

Tests:

• Added test_get_tools_metadata_has_server_name covering both tool_name_prefix=True and tool_name_prefix=False.
• Updated test_load_mcp_tools_with_annotations to include the new key in its expected dict.

63 passed, 2 skipped locally.