Add mixed precision support

[julienperichon]

Context

Some Sicara project using Keras FSL was needing support for mixed precision, which was not possible for some classes until now.

Changes

• Updated ClippedBinaryCrossentropy loss to cast float parameters to loss dtype when computations are done in type other than float32
• Update LearntNorms to use float32 precision in last activation layer when using mixed precision

To Do after this

• Update all repository for mixed precision support

[codecov]

Codecov Report

Merging #46 (4c67235) into master (149dd31) will increase coverage by 0.65%.
The diff coverage is 100.00%.

@@            Coverage Diff             @@
##           master      #46      +/-   ##
==========================================
+ Coverage   65.93%   66.59%   +0.65%     
==========================================
  Files          44       44              
  Lines         963      982      +19     
==========================================
+ Hits          635      654      +19     
  Misses        328      328              

Impacted Files	Coverage Δ
keras_fsl/losses/gram_matrix_losses.py	100.00% <100.00%> (ø)
keras_fsl/losses/tests/gram_matrix_losses_test.py	100.00% <100.00%> (ø)
keras_fsl/models/head_models/learnt_norms.py	100.00% <100.00%> (ø)
..._fsl/models/head_models/tests/learnt_norms_test.py	96.15% <100.00%> (+1.41%)	⬆️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 149dd31...4c67235. Read the comment docs.

[lmontier]

Peux tu mettre le lien vers la branch où le calcul de la mixed precision a fonctionné?