Windows support: replace the SIGALRM per-call timeout

[LuShadowX]

The per-call timeout in src/nodrift/replay.py uses signal.SIGALRM, which does not exist on Windows. The whole tool is therefore Unix-only.

What to do

Replace the alarm with something portable. Options worth weighing:

• run each call in a worker thread with a join timeout (cannot kill a hung call, but does not block the run)
• a watchdog thread that raises in the main thread via ctypes
• run replay batches in a subprocess with a timeout

A mutation can easily produce an infinite loop, so the timeout is not optional — that is the constraint to design around.

Relevant file: src/nodrift/replay.py

[LuShadowX]

Closed by #20. The timeout became portable in #16; #20 fixed the other half — POSIX-only temp-path scrubbing — and added windows-latest to the CI matrix. All three Windows Python versions now pass, so the claim is backed by CI rather than asserted.