fix: pipe stderr asynchronously to support Jupyter notebook environments

[michaelxer]

Summary

Fixes stderr handling in stdio_client to work correctly in Jupyter notebook environments by always piping stderr and reading it asynchronously, rather than passing sys.stderr directly as a subprocess file descriptor.

Fixes #156

Problem

In Jupyter notebook environments, sys.stderr is not a regular file descriptor — it's a rich text widget that cannot be passed to subprocess.Popen or anyio.open_process. This causes stdio_client to fail with an error when trying to start MCP servers from within Jupyter notebooks.

Solution

Instead of forwarding sys.stderr (or a custom errlog stream) directly to the subprocess, we now:

1. Always pipe stderr (subprocess.PIPE) — both on Unix and Windows
2. Read stderr asynchronously in a new stderr_reader task that runs alongside stdout_reader and stdin_writer
3. Detect Jupyter environments via _is_jupyter_notebook() which uses isinstance(get_ipython(), ZMQInteractiveShell) to check for ZMQ-based IPython kernels
4. Output with appropriate formatting — ANSI red coloring via print() in Jupyter, plain text via the errlog stream elsewhere

Changes

• src/mcp/client/stdio.py:

  • Added _is_jupyter_notebook() helper function
  • Added stderr_reader() async task in stdio_client()
  • Removed errlog parameter from _create_platform_compatible_process()
  • Always set stderr=subprocess.PIPE in process creation

• src/mcp/os/win32/utilities.py:

  • Removed errlog parameter usage, always use stderr=subprocess.PIPE

• tests/issues/test_1027_win_unreachable_cleanup.py:

  • Updated call to _create_platform_compatible_process() to match new signature (removed errlog arg)

Testing

• All existing tests pass
• Pyright type checking passes with 0 errors
• Ruff linting passes