Increase VAE decode memory estimates

[RyanJDick]

Summary

This PR increases the working memory estimates for VAE decode operations. The previous values were set based on memory allocations and were pretty aggressive about trying to make full use of the available VRAM. The updated values were set based on experimentally-observed reserved memory (not just allocated) levels, and also aim to be conservative by including some buffer room.

This change is intended to address reports of slow VAE decoding.

The more conservative memory estimate values could cause more memory to be offloaded from the GPU during VAE decode. This in turn could result in slower model reloading on subsequent runs. The net impact of this change is expected to be positive, but there may be a noticeable regression for some users. We will want to keep a close eye on this when it is released in an RC.

Related Issues / Discussions

N/A

QA Instructions

I tested several common VAE decode scenarios with memory profiling enabled. Results:

All tests run with `pytorch_cuda_alloc_conf: "backend:cudaMallocAsync"`.

fp32, SD1, 512x512
- Estimated: 2450 MB
- Reserved:  1696 MB

fp16, SDXL, 1024x1024
- Estimated: 4400 MB
- Reserved:  3232 MB

fp32, SDXL, 1024x1024
- Estimated: 9050 MB
- Reserved:  5568 MB
Note: Even in fp32, we apply some optimizations to still run some layers in fp16. We don't account for this in the estimate, which explains why there is so much buffer in this case.

fp16, FLUX, 1024x1024
- Estimated: 4400 MB
- Reserved:  2944 MB

fp16, SD3, 1024x1024
- Estimated: 4400 MB
- Reserved:  3136 MB

Merge Plan

• Merge Add pytorch_cuda_alloc_conf config to tune VRAM memory allocation #7673 first and change target branch to main
• Monitor this change closely during RC. The net benefit is expected to be positive, but there is a risk of regression on some systems.

Checklist

• The PR has a short but descriptive title, suitable for a changelog
• Tests added / updated (if applicable)
• Documentation added / updated (if applicable)
• Updated What's New copy (if doing a release after this PR)