fix: file-log rotation never fires again after process start

[raajkumars]

Summary

The per-process file log's rotation trigger, introduced in #1438, is a generator that returns True exactly once and False forever after (_once_then_never). loguru calls the rotation callable on every message, so this correctly rotates once on startup — giving a fresh log file each run — but then silently disables every subsequent rotation for the rest of the process's life. The retention/compression settings on the same logger.add() call imply ongoing rotation is happening; it never is, so the log file grows without bound on a long-running process.

Fix

Replace the generator with a stateful rotation policy (_rotation_policy): still force-rotate on the first message (preserving the original fresh-log-per-run intent), then rotate again once the file crosses 200MB — the check the generator never actually performed despite the retention/compression config being set up as if it had. The threshold is currently hardcoded (_ROTATION_MAX_BYTES), not exposed as a CLI flag or env var.

Tests

Added src/exo/shared/tests/test_logging_rotation.py, 5 tests against the new policy directly (no I/O, a minimal fake file object standing in for what loguru passes to the rotation callable):

• first call always rotates regardless of size
• second call does not rotate while under the size threshold
• rotates once file size + message length crosses the threshold
• only the first call ever forces a rotation (no repeated startup-style behavior)
• a tell() raising OSError is handled without propagating

All pass. ruff check, ruff format --check, and basedpyright are clean on both changed files.