Possibly replace tensorflow usage in data_utils with lighter-weight tool

[cisaacstern]

The tensorflow dependency in climsim_utils.data_utils is pretty heavy, and among other things makes using that module on Mac M1 a bit cumbersome. In that module, tensorflow is actually only used for data loading:

ClimSim/climsim_utils/data_utils.py

Lines 312 to 316 in f7d092c

	return tf.data.Dataset.from_generator(
	gen,
	output_types = (tf.float64, tf.float64),
	output_shapes = ((None,124),(None,128))
	)

Per @jerrylin96, we may be able to use another approach here, which would be desirable if possible.

[jerrylin96]

I think we need to get rid of tensorflow usage entirely in data_utils.py. This will result in significant streamlining. I propose refactoring the creation of the numpy arrays using Apache Spark instead of the Tensorflow dataloader. Using that for preprocessing is a vestige of some really crude code macgyvering that doesn't belong in a polished repo.