Load numpy first if available

[jrapin]

Torch users often get the following issue:

Error: mkl-service + Intel(R) MKL: MKL_THREADING_LAYER=INTEL is incompatible with libgomp.so.1 library.
        Try to import numpy first or set the threading layer accordingly. Set MKL_SERVICE_FORCE_INTEL to force it.
srun: error: learnfair5096: task 0: Exited with exit code 1
srun: Terminating job step 26303611.0

This is due to numpy and torch being loaded in a different order when reloading the pickle. We need a way to make sure numpy is loaded first instead.

[jrapin]

Other options:

• avoid loading if some env variable is present
• look in an env variable for packages to load dynamically ahead of unpickling

In any case, the current patch does work in my case.

[gwenzek]

For now we only had issues with numpy and torch import order, this should fix it.
I think in the general case we could dump sys.modules.keys() which is the list of module in the order in which they have been imported. This would be simillar to #1596 .

[jrapin]

For now we only had issues with numpy and torch import order, this should fix it.
I think in the general case we could dump sys.modules.keys() which is the list of module in the order in which they have been imported. This would be simillar to #1596 .

I am a bit afraid that could become black magic, but we could try eventually if we find a new problem. Should we merge this in the meantime then?