FID metric is returning NaNs

[JessyD]

As shown in this notebook the FID returns NaN when given the synthetic and real images for the Hand dataset from MedNIST. However, as shown in this other notebook, the FID returns a value when using a test array.

[JessyD]

The _sqrtm_newton_schulz function is returning NaNs after a few iterations. Therefore, I will adapt the code for the FID metric using a similar implementation as Pytorch, as they do not use the Newton-Schulz optimization to see if the NaNs persist.

[Warvito]

After adding .double() it is working as expected when conputing the FID of the mimic model https://github.com/Warvito/generative_mimic/blob/main/src/python/testing/compute_fid.py

[Warvito]

Since there is a discrepancy between the iterative approach and the original one, it will be better to adopt the original one