Loftusa patch 3

[loftusa]

@bdpedigo proposed fix to #1035. Kinda janky, but it bought my imports down to 2.35 seconds:

[bdpedigo]

linear_sum_transport gets called many many times during an optimization so (if i understand what this will do correctly) that import would also happen over and over again if that code is called - not sure that is viable. would ideally get at something closer to the root cause of this (for instance, why does POT need to even import tensorflow here...)

[bdpedigo]

closing since there hasnt been action on this but happy to revisit if there's interest