feat: Add more powerful drift windowing strategies, warmup and dynamic thresholds

[robinholzi]

Motivation

Allow for more powerful drift windowing, dynamic drift threshold as well as a warmup period (for calibration)

[github-actions]

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

[codecov]

Codecov Report

Attention: Patch coverage is 87.82895% with 37 lines in your changes missing coverage. Please review.

Project coverage is 84.86%. Comparing base (3de6590) to head (9594453).

Files	Patch %	Lines
.../internal/triggers/drift/detection_window/time_.py	50.00%	18 Missing ⚠️
...n/supervisor/internal/triggers/datadrifttrigger.py	86.36%	12 Missing ⚠️
...internal/triggers/drift/detection_window/amount.py	90.00%	3 Missing ⚠️
modyn/supervisor/internal/triggers/utils/utils.py	66.66%	2 Missing ⚠️
...dyn/config/schema/pipeline/trigger/drift/config.py	91.66%	1 Missing ⚠️
...pervisor/internal/triggers/drift/detector/alibi.py	83.33%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #564      +/-   ##
==========================================
- Coverage   84.87%   84.86%   -0.01%     
==========================================
  Files         224      233       +9     
  Lines       10300    10480     +180     
==========================================
+ Hits         8742     8894     +152     
- Misses       1558     1586      +28     

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

[robinholzi]

@MaxiBoether fyi. this should be ready now :)
** I haven't run a pipeline with this yet, so maybe we still need some minor tweaks in a followup PR.

[MaxiBoether]

Thank you, Robin. I did a first pass. While I did not fully grasp everything yet, I think my major feedback for the next iteration of this PR is that we should think about the interactions between the different components and potentially document them in some README.

In particular, we have aggregation logic for decisions in the pydantic module, we have the drift detectors and then some duplicated decision logic in the decision engine. I think we want to consolidate the how we make drift/no drift decisions in a central place somehow. I also propose to reconsider the naming of the windowing manager. Last, I also encountered some duplication in the new pydantic classes.

I hope my comments make sense. Basically I think the naming and levels of abstraction need some adjustment. Happy to discuss tomorrow or via Github comments!

[MaxiBoether]

Thank you, Robin! This is looking good. Most of my comments are nitpicky, I had a couple of smaller questions in some parts

[MaxiBoether]

Thank you!! Some last nitpicky comments, feel free to merge afterwards.