Arbitrary prediction time transforms

[ziw-liu]

Use a standalone module to perform arbitrary forward and inverse transforms at prediction time. This is also the first step of moving to dependency injection base architecture configuration for our lightning modules.

Below is an example config field to perform a 2x XY zoom with linear interpolation before inference, and a 0.5 XY zoom with area interpolation (no aliasing) with afterwards to restore the image size. This is roughly what @tayllatheodoro has been doing manually for the virtual staining of OPS data (cc @ieivanov).

model:
  class_path: viscy.translation.engine.AugmentedPredictionVSUNet
  init_args:
    model:
      class_path: viscy.unet.networks.fcmae.FullyConvolutionalMAE
      init_args:
        in_channels: 1
        out_channels: 2
        encoder_blocks: [3, 3, 9, 3]
        dims: [96, 192, 384, 768]
        decoder_conv_blocks: 2
        stem_kernel_size: [5, 4, 4]
        in_stack_depth: 15
        pretraining: false
    forward_transforms:
      - class_path: viscy.transforms.BatchedZoom
        init_args:
          scale_factor: [1, 2, 2]
          mode: trilinear
    inverse_transforms:
      - class_path: viscy.transforms.BatchedZoom
        init_args:
          scale_factor: [1, 0.5, 0.5]
          mode: area

[ziw-liu]

@edyoshikun I'm merging this to simplify workflows. Please review in #195.