Adding another way of calculating gradients in downsamplers

[XianzheMa]

This PR addresses #505.
A flag is introduced, to switch between using purely the last layer gradients, or the concatenation of last layer gradients and penultimate layer gradients, in selection algorithms.

[github-actions]

^{( % to main)
( % to main)}

[codecov]

Codecov Report

Attention: Patch coverage is 91.54930% with 6 lines in your changes missing coverage. Please review.

Project coverage is 82.83%. Comparing base (70ce04c) to head (84c20ff).

Files	Patch %	Lines
...sampling_strategies/craig_downsampling_strategy.py	0.00%	2 Missing ⚠️
...ling_strategies/gradmatch_downsampling_strategy.py	0.00%	2 Missing ⚠️
...ing_strategies/submodular_downsampling_strategy.py	0.00%	2 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #540      +/-   ##
==========================================
- Coverage   82.84%   82.83%   -0.01%     
==========================================
  Files         220      220              
  Lines       10202    10232      +30     
==========================================
+ Hits         8452     8476      +24     
- Misses       1750     1756       +6     

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

[MaxiBoether]

Thanks, my only concern is (and I think codecov complains as well) that we should at least test whether the new code path does not crash or something. But the changes itself look good and I tried to validate that the old logic stay the same :)

[MaxiBoether]

Thanks, the new tests look good. Feel free to merge!