Low memory inference fix

[ghost]

Resolves two separate issues:

1. Use spawned workers instead of forked workers to resolve CUDA_NOT_INITIALIZED_ERROR (closes Cuda Error with low_mem #36)
2. Passes seed to low_mem inference correctly, and in a way that does not require tensorflow to be pickled (closes NameError: name 'self' is not defined message appears when running in --low mem option #491, can't pickle tensorflow #529, Low memory inference fix #536)

Requesting @mineLdiver to download the 491_fix_lowmem_seed branch of my fork and test these changes to make sure they work.

[ghost]

This change has been tested by myself and @mineLdiver (#535 (comment)).

[ghost]

Thanks @CorentinJ . We can delete the low_mem_fix branch since it is now in master.