speed up and reduce VRAM of QWEN VAE and WAN (less so)

[rattus128]

This is a fast path for QWEN VAE by taking advantage of the fact that zero padding a 3d convolution on a single frame is mathematically equivalent to just slicing the weight.

I dont have a good explanation for why it is this much faster. I was expecting VRAM savings and hoping for some speed, but I got less VRAM and a ton of speedup. pytorch works in mysterious ways.

Example test conditions:

RTX5090

QWEN VAE Encode -> Decode (3840x2160)

Before:

Requested to load WanVAE
loaded completely; 7324.16 MB usable, 242.03 MB loaded, full load: True
0 models unloaded.
Unloaded partially: 242.03 MB freed, 0.00 MB remains loaded, 22.78 MB buffer reserved, lowvram patches: 0
Prompt executed in 36.82 seconds

After:

Requested to load WanVAE
loaded completely; 7324.16 MB usable, 242.03 MB loaded, full load: True
0 models unloaded.
Unloaded partially: 242.03 MB freed, 0.00 MB remains loaded, 22.78 MB buffer reserved, lowvram patches: 0
Prompt executed in 2.42 seconds

WAN2.2 VAE Encode -> Decode (1920x1088x81f)

Before (31.8GB):

Requested to load WanVAE
loaded completely; 7148.38 MB usable, 242.03 MB loaded, full load: True
Prompt executed in 51.00 seconds

After(30.8GB):

Requested to load WanVAE
loaded completely; 7148.38 MB usable, 242.03 MB loaded, full load: True
Prompt executed in 41.56 seconds

This WAN test point is probably thrashing the pytorch and cuda-malloc-async allocators to stay on the VRAM ceiling which is why its more of a speedup than a VRAM saving.