Add ground truths only data loader

[fsherry]

This adds a CTBenchmarkingExperiment, called GroundTruthCT, and a corresponding task 'groundtruth', which only loads the ground truth reconstructions, for example for use in training a denoiser. The time saved processing the sinograms (which happens on CPU) allows for higher GPU utilisation.

The issue of time wasted on preprocessing sinograms when only training denoisers was mentioned in #92.

[AnderBiguri]

Happy to to these changes myself, just noting them so I don't forget :D

[fsherry]

@AnderBiguri Please check that it was right to add this, but it seemed like there was a missing else in the branching on self.do_recon for target (which seems to be ok for input). As a result, it would compute a reconstruction, but then still load the reconstruction from file, overwriting the computed reconstruction. The relevant snippet is here, it looks like there should be an else before line 419, which I added in commit e97a008:

LION/LION/data_loaders/deteCT.py

Lines 408 to 421 in 87a962c

	# if target is reconstruction, we need to load the reconstruction
	elif self.task in ["recon2recon", "sino2recon", "groundtruth"]:
	if self.do_recon:
	sinogram = self.__load_and_preprocess_sinogram__(
	index, self.target_mode
	)
	op = deteCT.get_operator()
	if self.recon_algo == "nag_ls":
	target = nag_ls(op, sinogram, 100, min_constraint=0)
	elif self.recon_algo == "fdk":
	target = fdk(op, sinogram)
	target = self.__load_and_preprocess_reconstruction__(
	index, self.target_mode
	)

[AnderBiguri]

@fsherry correct, good catch!