fix: warm latent cache loading does an unnecessary model swap

[Mandrakia]

Problem

LatentCachingMixin wraps unconditionally the caching loop with

set_device_state_preset('cache_latents')
… 
restore_device_state()

On a warm start where every latent is already cached, the loop body is just os.path.exists checks — nothing is encoded — yet the device-state preset still moves models (notably the 9B transformer) off the GPU and back on, once per dataset × resolution pass. That move, not the encoding, is what makes warm starts take 10s+.

Fix

Change device state only the first time we actually hit a cache miss (gated behind a did_move flag), and restore only if we moved. This mirrors the pattern cache_text_embeddings already uses for cache_text_encoder.

Cold starts are unchanged; warm starts skip the unnecessary GPU shuffle.