Optimize AbelianBackend split_legs

[Jakob-Unfried]

From comment that was in the code:

# TODO (JH) below, we implement it by first generating the block_inds of the splitted tensor and
# then extract subblocks from the original one.
# Why not go the other way around and implement
# block_views = self.block_backend.block_split(block, block_sizes, axis) similar as np.array_split()
# and call that for each axis to be split for each block?
# we do a lot of numpy index tricks below, but does that really save time for splitting?
# block_split should just be views anyways, not data copies?