Pull memory improvements from neonsecret's fork

[SkylarKelty]

See basujindal@47f8784

This tends to allow about +64px in either direction within the same GPU memory and no change in image output

[AscendedGravity]

Ran this PR against the current main branch commit and compared. Running on a GTX 1070 8gb. Using optimized_turbo.

The following prompt was used:

landscape by Thomas Kinkade seed:2464003812 width:896 height:512 steps:50 cfg_scale:7.5 sampler:k_lms

Results:

Main Branch - Took 78.21s total (78.21s per image) Peak memory usage: 6675 MiB / 8192 MiB / 81.478%

This PR - Took 77.45s total (77.45s per image) Peak memory usage: 5941 MiB / 8192 MiB / 72.521%

Reduced memory footprint. No discernable change in image generated.

If PR is merged, may want to consider increasing the max dimensions slider in the UI since this makes generations above 1024 (at least on one side) more obtainable. Although, after 1024px, generations can get a bit weird.