feat: save and reuse trained shadow models

[rpreen]

Resolves #347 and resolves #351

• Move shadow model training to sacroml/attacks/utils.py and save trained models and indices to subfolders;
• Reuse any previously trained models;
  • This means that if different LiRA modes need to be computed (e.g., offline and online) a new attack object can simply be created with a different mode and it will quickly run the attack using previously trained models and save the output to the report JSON and PDF.
  • Any other future attacks that may use shadow models can reuse these trained models;
  • A small number of shadow models (e.g., 10) can be initially be used to compute scores and then if needed a subsequent attack object with a larger number (e.g., 20) can be run and it will only need to train the additional models;

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 99.72%. Comparing base (f97195f) to head (2d713c3).

Additional details and impacted files

@@           Coverage Diff           @@
##             main     #357   +/-   ##
=======================================
  Coverage   99.72%   99.72%           
=======================================
  Files          43       44    +1     
  Lines        4343     4386   +43     
=======================================
+ Hits         4331     4374   +43     
  Misses         12       12           

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

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.