Cache the Wan transformer expert(s) between consecutive trainings against the
same base (keyed by base_path+quantize) so a back-to-back job skips the very slow
reload (tens of minutes for A14B). Only this job's adapter + gradient-checkpoint
hooks are removed at teardown; the base transformer(s) stay resident. Since 4-bit
weights can't move to CPU, they hold GPU VRAM between jobs — so the external VRAM
releaser now drops the Wan cache too when a generation needs the GPU, and the
error path clears both caches.

Also report training progress every step (cheap dict update) instead of every
10, so the web UI bar advances smoothly once steps begin.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>