Support multiple MCP servers in MCPEnvironment

[taivu1998]

Summary

This PR adds multi-server MCP support to MCPEnvironment so a single environment can route tool calls across multiple named MCP servers.

Fixes #321.

What changed

• Added mcp_servers support to MCPEnvironment for configuring multiple named MCP servers.
• Added tool_name_to_server_name routing for disambiguating tool names when more than one server exposes the same public tool.
• Kept MCPConnectionManager single-server and moved the multi-server logic into MCPEnvironment as a lightweight registry/router layer.
• Preserved the legacy single-server mcp_server_command / mcp_server_args / mcp_server_env API for backward compatibility.
• Updated tool-call execution to handle malformed tool calls more gracefully and to support both dict and JSON-string arguments.
• Preserved original tool-call ordering in tool_outputs even when calls are routed to different MCP servers.
• Added cleanup/rollback behavior so failed startup or routing validation does not leave partially initialized global MCP managers behind.
• Added MCP example documentation showing the new multi-server API.

Design notes

• Duplicate public tool names now fail fast unless explicitly mapped with tool_name_to_server_name.
• Duplicate detection uses the effective public tool names exposed today, including underscore aliases for tools whose MCP names contain hyphens.
• Single-server behavior remains permissive: if exactly one server is configured, unresolved tool names still fall back to that server so existing behavior is preserved.
• Partial per-server failures are isolated so successful tool outputs from other servers are still returned.

Testing

• /Users/vuductai/Documents/Projects/rllm/.venv/bin/python -m ruff check rllm/environments/tools/mcp_env.py tests/envs/test_mcp_env.py
• /Users/vuductai/Documents/Projects/rllm/.venv/bin/python -m pytest tests/envs/test_mcp_env.py tests/agents/test_tool_agent.py tests/envs/test_tool_env.py -q

Coverage added

• Multi-server initialization
• Explicit routing for ambiguous tool names
• Duplicate-name validation
• Cross-server tool execution routing
• Partial server failure handling
• Synthetic ID stability for missing tool-call IDs
• Interleaved output ordering across servers
• Startup and routing-validation rollback behavior
• from_dict support and non-mutation of input args