fix: cap tiled VAE decode output at 8GB RAM, disk-back above that, an…

[Juan Fidel (johnnycaifan)]

…d fix a non-in-place normalization

Both decode() and decode_spatial_temporal() preallocate their full output tensor (and decode()'s weight tensor) via torch.zeros/torch.empty sized for the entire batch (batch x frames x height x width x channels), with no cap. For a long/high-res video batch this can require tens of GB in a single allocation - on a 64GB system, running this repeatedly pushed system RAM past 97% used and required killing the process to avoid a full freeze.

Adds _allocate_decode_output(), matching the same threshold shape as the RTX Upscaler & Refiner fix in DaSiWa-Nodes: allocate normally under 8GB, back the tensor with torch.from_file() (checking real free disk space first) between 8GB and 64GB, and refuse cleanly above that instead of risking an OOM freeze.

Also fixes a separate leak in decode_spatial_temporal(): the normalization step

output /= weights + 1e-8

looks in-place but isn't - weights + 1e-8 allocates a brand-new full-size tensor before the division even starts, via the plain allocator, bypassing the disk-backed buffer above. Changed to

weights.add_(1e-8)
output /= weights

which is truly in-place. On a real render this second, unprotected allocation was still enough on its own to push RAM past the same 97% threshold even with the main output buffer disk-backed.