Vae decode oom feature request

### Feature Idea

When regular VAE decoding OOMs, could ComfyUI first unload/release other models and retry regular VAE decoding before falling back to tiled decoding? On VRAM-constrained systems, this may allow normal decoding to succeed and avoid the performance cost of tiled decoding.

### Existing Solutions

_No response_

### Other

_No response_

The issue names no files, tests, or entry points. Start by locating the VAE decode path and its OOM handling and tiled-decoding fallback, then inspect how other models are released; done means regular decoding is retried after release before tiled decoding is used, with coverage for the constrained-memory case.

Written by the indexing model from the issue text.