fix: wrap MCP call_tool coroutine in a Task so CancellationToken can cancel it

[Pujitha Paladugu (pujitha24)]

Why are these changes needed?

McpWorkbench.call_tool links a CancellationToken to the object returned by McpSessionActor.call("call_tool", ...). That object was a bare coroutine (McpSessionActor._run_actor did result = session.call_tool(...) without ever wrapping it in a Task/Future), but CancellationToken.link_future calls future.cancel(), which a coroutine object does not implement. As a result:

• Cancelling the token while a tool call is in flight raised AttributeError: 'coroutine' object has no attribute 'cancel' out of CancellationToken.cancel(), and the in-flight MCP tool call was never actually cancelled — it ran to completion anyway.
• Cancelling the token before calling call_tool raised the same AttributeError inside call_tool's try block, which was swallowed by the generic except Exception handler and surfaced to the caller as a confusing is_error=True result containing the AttributeError string instead of a clean cancellation signal.

The sibling implementation McpToolAdapter._run (same package, _base.py) already does this correctly by wrapping the coroutine with asyncio.ensure_future(...) before linking the cancellation token, and by checking cancellation_token.is_cancelled() up front.

Approach

• _actor.py: in McpSessionActor._run_actor's call_tool branch, wrap session.call_tool(...) in asyncio.ensure_future(...) before handing it back, so the object linked to the CancellationToken is an actual Task with a working .cancel().
• _workbench.py: in McpWorkbench.call_tool, check cancellation_token.is_cancelled() before dispatching to the actor and raise asyncio.CancelledError immediately, mirroring McpToolAdapter._run's existing pattern, instead of starting a doomed tool call. asyncio.CancelledError is a BaseException (Python 3.8+), so it is not caught by the existing except Exception clause and propagates to the caller, matching the already-shipped behavior of McpToolAdapter._run (see existing tests test_mcp_tool_adapter_run_cancelled_before_call / _during_call) and of StaticWorkbench.call_tool (autogen_core), which also only catches except Exception.
• The other _run_actor command branches (list_tools, list_prompts, list_resources, list_resource_templates, read_resource, get_prompt) are left unchanged: none of them accept a cancellation_token in McpWorkbench and none call link_future, so they cannot trigger this bug.

Validation

Added python/packages/autogen-ext/tests/tools/test_mcp_cancellation.py with two targeted regression tests, verified to FAIL with the exact reported AttributeError before this fix (confirmed via git stash on the two source files) and PASS after:

• test_call_tool_cancellation_stops_inflight_call: starts a mocked slow call_tool, cancels the token mid-flight, asserts token.cancel() does not raise and the call is cancelled promptly (< 1s) instead of running to completion (mocked 5s sleep).
• test_call_tool_pre_cancelled_token_reports_cancellation: cancels the token before calling call_tool, asserts asyncio.CancelledError is raised and the underlying session's call_tool is never invoked.

Commands run (all passing) from python/, in the uv-managed virtualenv per python/README.md:

• poe format (autogen-ext) — 1 file reformatted (the new test file), no other changes.
• poe lint (autogen-ext) — all checks passed.
• poe mypy (autogen-ext) — success, no issues in 198 source files.
• poe pyright (autogen-ext) — 0 errors, 0 warnings.
• pytest packages/autogen-ext/tests/tools/test_mcp_cancellation.py packages/autogen-ext/tests/tools/test_mcp_actor.py packages/autogen-ext/tests/tools/test_mcp_workbench_warnings_and_errors.py packages/autogen-ext/tests/tools/test_mcp_tools.py — 83 passed, 1 skipped (pre-existing GitHub-token-gated skip, unrelated).
• Full tests/tools -k mcp — 147 passed, 1 skipped.

I also ran the full autogen-ext package test suite once; it showed unrelated failures (docker, playwright, llama_cpp model download, http_tool) that I traced to the sandbox's local disk being 99% full (166Mi free) — re-running the same MCP-adjacent tests without parallelism, on their own, passed cleanly, confirming those were environmental flakes and not caused by this change.

Related issue number

Closes #8265

Checks

• I've included any doc changes needed for https://microsoft.github.io/autogen/. See https://github.com/microsoft/autogen/blob/main/CONTRIBUTING.md to build and test documentation locally.
• I've added tests (if relevant) corresponding to the changes introduced in this PR.
• I've made sure all auto checks have passed.