GPU accelerated dead leaves generator

[Jamy-L]

The current generator is currently very slow, as it requires to compute a high resolution image and downsample it by a factor of up to 5.

A solution is to implement that using numba kernels to use the GPU.

To do:

• [x] Generate the HR image using numba device arrays
• [x] Convert to pytorch and apply anti-aliased downsample
• [x] Convert numba to pytorch gpu tensors on the fly

[Jamy-L]

The new dataset is working but not connected yet. The speedup is roughly x2, and the memory cost is very low.

[balthazarneveu]

• [x] add pytest

[Jamy-L]

• Pytest added for the new dataset, plus a new test with a short gpu vs cpu benchmark. These tests are skipped if cuda is not installed (numba.cuda.is_available()).

• Fixed a bug for ds=1 that crashed the GPU dataloader !

• The dataloader in grey mode will now output shape [h, w, 1] instead of shape [h, w, 3] (duplicated channel). The visualisation may need adjustments