Add VRAM free after sample images

[v0xie]

This PR addresses an issue with VRAM usage after generating sample images.

Without using this PR, when training with SDXL and batch size > 1, the first time during training when sample images are generated, I would run out of VRAM even with 24GB (no --xformers or --mem_eff_attn).

Notably, the sample_images_common method in train_util calls clean_memory_on_device near the end of the function, but I hypothesize that some of the memory is still used after the operation at the end. So calling it again after exiting the function clears the remaining.

I'm not sure if anyone else has had this issue or if it's an issue with my configuration only. Please feel free to close this PR if it's unnecessary. Thank you for your consideration.

[araleza]

This seems good, I think that at times in the past, I've seen training work up until the sample images were generated, and then not be able to restart afterwards.

[kohya-ss]

Thank you for this!

Notably, the sample_images_common method in train_util calls clean_memory_on_device near the end of the function, but I hypothesize that some of the memory is still used after the operation at the end. So calling it again after exiting the function clears the remaining.

This makes sense... I'm working on SD3 now, so I will merge this sooner when I finish SD3.