bumped up to python3.12 & tf 2.16 (some tf code changed were needed)

[nethemis]

Bumping up to newer python/TF versions. Should help run the model marginally faster on latest MacOS CPUs/GPUs.

Summary of changes:

• bumped up the library versions
• replaced direct Keras calls with indirect ones through TF.
• replaced K.mean() with tf.reduce_mean() which is the suggested way.

Validation steps taken:

• NN training results stay the same
• Unit tests pass

[nethemis]

Hi @MichaelVDeSanti, @moselhy,

The CI checks are failing for this PR of mine. I do not have access to the logs to understand the reason why, since it requires University access. Could either of you kindly have a look?

Neofytos