Implement MCP Tasks runtime for .LongRunning() commands (ModelContextProtocol.Extensions.Tasks)

[carldebilly]

Context

SDK 2.0 extracted MCP Tasks into ModelContextProtocol.Extensions.Tasks (2.0.0-preview.3): IMcpTaskStore/InMemoryMcpTaskStore, task results/notifications (CreateTaskResult, GetTaskResult, CancelTaskResult, completed/failed/cancelled/input-required), and client helpers (CallToolAsTaskAsync, CallToolWithPollingAsync). The per-tool Tool.Execution augmentation is gone from the protocol surface.

Since PR #71, Repl keeps .LongRunning() in its own model (help/docs) and deliberately does not advertise task support at the protocol level — per #51's non-goal: no advertisement until polling/cancel/update works end-to-end.

Scope

• Reference ModelContextProtocol.Extensions.Tasks and wire a task store (in-memory by default, pluggable).
• Route .LongRunning() tool calls through the task lifecycle: create → progress → completed/failed/cancelled, honoring cancellation.
• Handle tasks/get, tasks/update, tasks/cancel and the corresponding notifications.
• Advertise task capability once the runtime is complete (whatever shape the SDK exposes at that point — the per-tool augmentation no longer exists).
• End-to-end tests using the SDK client's CallToolAsTaskAsync/CallToolWithPollingAsync.
• Update docs/mcp-reference.md#sdk-and-protocol-versions and the .LongRunning() doc mentions when advertisement returns.

Notes

• Track SDK previews: the surface may still move before 2.0.0 stable.
• Parent epic: Ready for ModelContextProtocol v2.0 #51.