Extension SDK: Allow mid-turn tool list rebuild after MCP enable/disable

### Describe the feature or problem you'd like to solve

When an extension calls rpc.mcp.enable() mid-turn, the MCP server connects successfully (confirmed via mcp_server_status_changed events), but its tools don't appear in the LLM's tool list until tools_changed_notice fires at the next turn. session.rpc.tools.initializeAndValidate() is callable from joinSession() extensions but is a no-op (returns {} instantly). There is no way for an extension to force the runtime to rebuild the LLM's active tool list mid-turn. Tested with 3 MCP servers (mail, workiq, s360-breeze) — none had tools available same-turn.

### Proposed solution

Expose an API that lets extensions trigger a mid-turn tool list rebuild. Options:

1. `session.rpc.tools.rebuild()` — new method that forces the runtime to rebuild and inject the tool list mid-turn
2. Make `initializeAndValidate()` functional for extension sessions via `joinSession()` (currently a no-op)
3. Allow extensions to programmatically trigger `tools_changed_notice` mid-turn

This would benefit any extension that dynamically loads/unloads MCP servers for token optimization. Currently users must waste a turn after every MCP load, degrading the experience.

Diagnostic evidence and test methodology: https://github.com/gacurtin_microsoft/constellation/issues/4

### Example prompts or workflows

1. User: "Check my S360 compliance" → Extension loads s360-breeze MCP → Extension calls tools.rebuild() → LLM immediately calls s360-breeze tools — all in one turn

2. User: "Switch to comms profile" → Extension enables teams + mail MCPs, disables others → calls tools.rebuild() once → LLM sees all new tools immediately

3. User: "Search my email for the budget report" → Extension loads mail MCP on demand → rebuild → LLM calls mail-SearchMessages — no wasted turn

Use case: Constellation (https://github.com/lowdrag84/constellation) — profile-based MCP manager saving 16K-33K tokens/turn by only loading needed MCPs. Everything works except this one-turn delay.

### Additional context

## Diagnostic Evidence

APIs tested from extension via joinSession():

| API | Status |
|-----|--------|
| session.rpc.tools.initializeAndValidate() | Callable but no-op — returns {} in ~29ms |
| session.rpc.tools (full namespace) | Only has: handlePendingToolCall, initializeAndValidate |
| session.on('mcp_server_status_changed') | Works — events fire with "connected" status |
| session.rpc.mcp.list() | Works — returns {servers: [...]} |
| tools_changed_notice | Only fires between turns, not triggerable |

## Live Test Results

| MCP Server | Connected? | Tools Same-Turn? | Tools Appeared Via |
|------------|-----------|-----------------|-------------------|
| mail | ✅ | ❌ | tools_changed_notice (next turn) |
| workiq | ✅ | ❌ | tools_changed_notice (~90s later) |
| s360-breeze | ✅ | ❌ | tools_changed_notice (~90s later) |

## Environment
- Copilot CLI v1.0.56-1
- Extension SDK: @github/copilot-sdk/extension via joinSession()
- OS: Windows 11
- Extension: Constellation (profile-based MCP manager)

Start with the joinSession() extension entry point and session.rpc.tools.initializeAndValidate(), then trace how mcp_server_status_changed and tools_changed_notice affect the active tool list. Compare the listed API options and live MCP results; done means an extension can rebuild and expose newly enabled MCP tools during the same turn.

Rédigé par le modèle d'indexation à partir du texte de l'issue.