Allow custom up and downsampling layers in U-net

[quantumjot]

When the U-net model is created, we should support custom up/downsampling layers and allow Conv2DTranspose etc to use the correct number of filters:

cellx/cellx/networks/unet.py

Lines 104 to 108 in 74d688a

	# set up the up/downsampling
	# TODO(arl): these may already be instantiated with custom params
	# TODO(arl): if using a transpose convolution, we need to set the number of filters
	self._downsampling = downsampling()
	self._upsampling = upsampling()