refactor(cli): extract epochs ConfigError handling into a shared helper and polish the message

[jeqcho]

Follow-up to #145, which was fixed by #146. Two PRs raced on that issue and each got something right; this issue collects what the merged one left on the table.

Current state after #146: _cmd_run and _cmd_eval_set each carry an inline try/except ConfigError around their replace(task, epochs=...) call, with the ConfigError import repeated mid-function in both.

Wanted:

• Extract the two blocks into a shared _apply_epochs_or_exit helper next to _resolve_or_exit, following the same raise SystemExit(str(exc)) from exc idiom. fix(cli): guided error for --epochs 0/negative instead of raw traceback #148 already had this shape, including moving replace into the helper and importing Task under TYPE_CHECKING; it is the reference.
• Polish the message while at it: --epochs: Epochs count must be >= 1, got 0 doubles up. The issue proposed --epochs must be >= 1, got 0. Keep the task attribution in the eval-set path.
• Consider pytest.raises(..., match=...) exact-message pinning in the epochs tests (again per fix(cli): guided error for --epochs 0/negative instead of raw traceback #148) rather than substring asserts.

Small and self-contained; a good second contribution for anyone who worked on #145.