Improve Camelyon16 benchmark

[xavier-owkin]

Fix the number of sampled tiles to 10 000 and ensure reproducibility by fixing torch RNG state

Results obtained with the fed_benchmark.py script:
New one

Old one

[jeandut]

We would need to update the results running fed_benchmark on the new Camelyon16 and updates corresponding results (leaving a copy of the old one for legacy purposes) in this PR as well as its plot putting old and new results side by side.

[jeandut]

Great work !
Can you:

• rename all old files with the suffix "_old" including (results_benchmark_fed_camelyon.csv to results_benchmark_fed_camelyon_old.csv so the file should not be changed directly only copied)
• new results are called like the old ones were before aka wo any suffix (no _vf)

[jeandut]

Great work ! Can you:

• rename all old files with the suffix "_old" including (results_benchmark_fed_camelyon.csv to results_benchmark_fed_camelyon_old.csv so the file should not be changed directly only copied)
• new results are called like the old ones were before aka wo any suffix (no _vf)

It seems I hallucinated on this one ! LGTM !