refactor: benchmark sanity checks

[gumityolcu]

Logging is done through Lightning logger determined through trainer configs. "wandb" or "tensorboard" can be given. The same logger is used to log training statistics, and the sanity check results in bench_prept/train.py script. I used "tensorboard" in the tests because wandb requires creating and exposing an API key. I did not use Hydra's instantiate since logger creation is handled through configs and Lightning.

Sanity checks are implemented as a seperate function that returns a dictionary of scores. The base class computes train and val accuracy and subclasses build on that. Added test for sanity check functionality.

[codecov]

Codecov Report

Attention: Patch coverage is 92.00000% with 4 lines in your changes missing coverage. Please review.

Project coverage is 91.07%. Comparing base (e02fda3) to head (ef89206).
Report is 9 commits behind head on dilya-bench-refactor.

Files with missing lines	Patch %	Lines
quanda/benchmarks/base.py	86.66%	2 Missing ⚠️
quanda/benchmarks/config_parser.py	77.77%	2 Missing ⚠️

Additional details and impacted files

@@                   Coverage Diff                    @@
##           dilya-bench-refactor     #271      +/-   ##
========================================================
+ Coverage                 90.64%   91.07%   +0.42%     
========================================================
  Files                        60       60              
  Lines                      2203     2253      +50     
========================================================
+ Hits                       1997     2052      +55     
+ Misses                      206      201       -5     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[dilyabareeva]

Thank you @gumityolcu for your hard work. I have:

• done some refactoring to isolate logging from the main library's functionality
• added tests for both wandb (set to offline) and tensorboard
• removed duplicated model-to-device placing