Lighten logger impact on installation and dynamic import

[matbun]

Summary

(Hopefully) solves this failing job, since PyPI does not support repositories as direct dependencies (e.g., prov4ml@git+https://github.com/matbun/ProvML@new-main).

Moreover, reduced the overhead of dynamic import of itwinai.loggers by moving the import of logging frameworks (e.g., mlflow) inside the dedicated itwinai wrapper. This also benefits all the code that depends on itwinai.loggers (e.g., itwinai.torch.trainer), reducing import time.

Also, moved the EmptyLogger at the bottom of itwinai.loggers module.

I also took the chance to simplify the imports in the itwinai.torch.trainer module.

From a simple experiment I ran on Vega's login node, simplifying imports reduced import time of TorchTrainer from ~26.6 to ~0.6 seconds:

start = time.time()
from itwinai.torch.trainer import TorchTrainer
print(f"time spent importing itwinai: {time.time() - start:.6f}")

I don't know how relevant/general this result is, but it seems that import time has been reduced.

---

Related issue :

Related to #283
Closes #278

[jarlsondre]

Looks god to me :)

Some might say divine, even

I thought of one more thing, sorry haha

[annaelisalappe]

Actually, I have one more general question, do you know generally the implications of moving imports into the classes? I wonder in cases such as ray train, where there would be multiple copies of the program running on each of the cores

[annaelisalappe]

Although it probably doesn't make a difference, because the classes are instantiated also only once per worker...

[jarlsondre]

My general understanding is that the import is only valid in the scope that it's declared, meaning that if you import it in a class or function, then it is only valid in that class or function. That being said, each copy should run "the entire program" anyway, except of course the parts that are hidden behind if-statements (rank == 0 etc.), so I wouldn't think that it's a problem. Not entirely sure of this, though.

[annaelisalappe]

Although it probably doesn't matter as the classes are each only instantiated once as well actually

[matbun]

Considering that each copy is a separate process, it should not interfere with inter-process operations. AFAIU, Python will try to dynamically import packages only once per session, so even if you are hiding your import inside multiple functions, it will actually get imported by Python only the first time and just reused the next times other functions import it.

[jarlsondre]

Looks divine to me