[Feature]: AI gateway — one MCP server for both mounts, reap shared-endpoint sessions, decouple UI tools from enable_mcp

[Nabil-Salah]

Requirement

As a Jaeger operator enabling the AI chat assistant,
I want the browser's per-turn UI tools to work over MCP without also exposing the
telemetry MCP server (ai.enable_mcp), the gateway's MCP endpoints to reap their
sessions at shutdown, and the shared and turn-scoped mounts to share one MCP server,
so that I can run a "chat only" assistant (UI tools, no external MCP exposure), MCP
sessions don't outlive the query server, and there is a single MCP server to trace
and reason about.

Problem

Part of #8890. These are RFC 0008 milestones M7.2 / M7.1 / M7 — three gaps in the
current gateway-hosted MCP server:

• M7.2 — UI tools coupled to ai.enable_mcp. The turn-scoped MCP endpoint
  (/api/ai/mcp/<mcpRouteID>/), which carries the per-turn UI tools, is mounted only
  when enable_mcp is set. UI tools also ride the ACP ext-method path today, so this
  is currently harmless — but M6 will drop that path, at which point enable_mcp: false with chat enabled would leave the agent no UI tools at all.

• M7.1 — the shared endpoint leaks sessions. mcptools.WrapHTTP returns a bare
  http.Handler, so the shared /api/ai/mcp/ mount has no closer and its MCP
  sessions outlive the query server. The turn-scoped mount got a closer in feat(ai-gateway): Announce the session-scoped MCP endpoint to the sidecar over HTTP #9009; the
  shared one never did.

• M7 — two MCP servers where one would do. The shared and turn-scoped mounts run
  two mcp.Server instances. The turn-scoped UI-tools middleware already degrades to
  telemetry-only when no turn is present, so one server can back both mounts.

Proposal

No response

Open questions

No response