Allow using ONNX compiled model in sklearn interface

[LeoGrin]

Fix #97

Downside compared to PyTorch

We don't do the memory estimation + reduction part, so it's easier to get memory error with large datasets (TODO)

(It seems that if if I export the onnx model on gpu I get less memory error afterward but not sure)

Quick speed analysis:

For 1Kx10 on a T4 GPU node (seems to be quite hardware dependent):

• first fit: 5s for ONNX (we initialize the session), 0.4s for PyTorch
• first predict: ~2s for ONNX, 6s for Pytorch
• 2nd fit: 0.05s for ONNX, 0.15s for PyTorch
• 2nd predict: 0.6-1.2s for ONNX (depending on where you export it, probably some more gains here?), 0.22s for PyTorch

Seems like for smaller datasets ONNX becomes faster even for 2nd predict.

Performance analysis on real datasets

Seemed good but need to rerun.

TODOs

• Make memory reduction work on ONNX
• Allow to set the onnx session from outside to work with sklearn cloning
• Check more carefully that we haven't degraded performance (e.g on categoricals)
• Can we make ONNX model faster by exporting it with the right dummy input? How much speed vary by hardware?