Fix Windows training: UTF-8 logs, direct plot spawns, safe checkpoint replace - #32

Summary

Three Windows-runtime failures, each hit on a live training run on this machine:

1. Log writers used the locale default encoding (cp1252 here). The model emits raw bytes, so a sample containing an unencodable character raised UnicodeEncodeError at sample time and killed the run mid-training. All four log writers are now explicit UTF-8, with errors="replace" on the session banner and sample log, where unencodable raw bytes are expected rather than exceptional.
2. refresh_plots() spawned a bash -c "... && ..." chain. On Windows, CreateProcess may resolve bash to WSL's bash.exe, which cannot execute D:/... paths - and stderr went to devnull, so every plot refresh failed silently. The tools are now spawned directly via sys.executable with argv lists. All children are tracked (a later spawn failing no longer loses earlier PIDs), and the skip-if-still-running guard is preserved; the two plots render concurrently, which is safe because they write distinct outputs.
3. Checkpoint saves publish with os.replace, which raises PermissionError (WinError 5) on Windows when any reader holds the destination - our own unwaited plot children open manifest.json and do exactly that. The window is milliseconds wide: one save survived, the next died. A bounded retry (8 attempts, exponential backoff, ~6.4 s worst case) keeps publication atomic; persistent failures still raise, and non-permission errors propagate immediately.

POSIX behavior is unchanged throughout.

Tests

tests/test_windows_runtime.py (10 tests, no torch import, no training, no production files): unicode sample-log writes, plot argv/process-state handling, retry exhaustion and immediate propagation of non-permission errors, plus a real Windows reader-lock race executed in a temp directory.

Untouched upstream sources reproduce these failures (5 failures + 3 errors, including the actual cp1252 UnicodeEncodeError and actual WinError 5); this branch passes all 10.

python -m unittest discover -s tests -v
# Ran 10 tests in 0.574s -- OK