Refactoring trainer to support to run prediction and unwatch on dq.finish

[franz101]

• Use Patch class
• Unwatch on dq.finish / config reset
• Be able to load trainer from checkpoint and run inference

[codecov-commenter]

Codecov Report

Merging #595 (ce89a72) into main (88a1422) will increase coverage by 0.07%.
The diff coverage is 91.57%.

@@            Coverage Diff             @@
##             main     #595      +/-   ##
==========================================
+ Coverage   84.88%   84.96%   +0.07%     
==========================================
  Files         161      161              
  Lines       12598    12688      +90     
==========================================
+ Hits        10694    10780      +86     
- Misses       1904     1908       +4     

Impacted Files	Coverage Δ
dataquality/utils/torch.py	80.42% <75.00%> (+0.42%)	⬆️
dataquality/utils/transformers.py	89.18% <88.33%> (-7.79%)	⬇️
dataquality/integrations/transformers_trainer.py	94.73% <97.29%> (+6.27%)	⬆️
dataquality/integrations/setfit.py	55.00% <100.00%> (-6.25%)	⬇️
dataquality/utils/patcher.py	100.00% <100.00%> (ø)
...sts/integrations/hf/test_text_classification_hf.py	100.00% <100.00%> (ø)
tests/start/test_start_auto.py	100.00% <100.00%> (ø)

📣 We’re building smart automated test selection to slash your CI/CD build times. Learn more

[Ben-Epstein]

perfect example of documentating the limitations of the integrations https://app.shortcut.com/galileo/story/5200/documentation-of-functionality-and-limitations-for-each-integration