Bugfix - Use Numpy array instead of mmap

[omer-dayan]

No description provided.

[noa-neria]

The only thing is that we add dependency on numpy, for creating a continuous buffer.
If we ever find another solution, it'd be better to remove it.

[noa-neria]

Can we import just the empty. i.e. from numpy import empty

[omer-dayan]

In python, importing partial package does not reduce import time (https://stackoverflow.com/questions/710551/use-import-module-or-from-module-import).

Its mainly convention.

I prefer keeping import numpy as np because it keep us with the context.
I mean comparing np.empty(... vs empty(...
In my opinion its more convenient to see the first as you immediately understand its numpy array.
In addition to that, it became a convention in the community to import numpy as np

For example, can you imagine people are doing from torch import zeros and using zeros(... instead of torch.zeros(... in the middle of the code?