Add masked loss

[kohya-ss]

Mask the loss value with conditioning image.

[Trojaner]

snip
(was related to deepspeed)

[Trojaner]

I cannot use reg images at the same time when conditioning_data_dir is defined, e.g.

[[datasets]]
caption_extension = ".txt"
resolution = [512,768]
  
  [[datasets.subsets]]
  image_dir = '/home/ml/checkpoints/regularization'
  is_reg = true
  class_tokens = "something"
  num_repeats = 1  

  [[datasets.subsets]]
  image_dir = '/home/ml/checkpoints/sd15/img'
  conditioning_data_dir = '/home/ml/checkpoints/sd15/mask'
  num_repeats = 10

voluptuous.error.MultipleInvalid: extra keys not allowed @ data['datasets'][0]['subsets'][1]['conditioning_data_dir']

[kohya-ss]

Unfortunately ControlNetSubset (used in the masked loss) doesn't support is_reg attribute. Is it fine to set num_repeats of each subset (reg and train) to make the total number of reg images (num_repeats * num of reg images) same as the total number of training images? I think it is very similar to use is_reg for reg images.