[tests] Centralize histogram production; Fix tests to only verify existing references and error if not all references are checked

Make code and tests flexible to the use of a pre-processed Montgomery dataset to speed-up testing where possible (closes #74)

Histograms comme be compared directly due to differences in how the
torch.resize transform is applied depending on the environment.
A statistical approach using pearson coefficients is used instead.
Testing of alexnet and densenet transforms have been temporarily
disabled as histograms or RGB images need to be handled differently.

Databases should now be available on the CI, we do not rely on a zipped
db anymore.

Samples are not square cropped during data loading anymore due to recent
changes. Cropping is model-dependent.

[tests] Revamp tests for montgomery, shenzhen, indian, and tbx11k; Erase tests for outdated RS modules

[tests] Simplifies test database unpacking by moving code to a pytest initializer; Make tests more unaware of the test database