Fix loading state dict for SingleTaskVariationalGP/ApproximateGPyTorchModel (#3251)

[esantorella]

Summary:

Context: See https://github.com/meta-pytorch/botorch/issues/3250 .

GPyTorchModel.load_state_dict previously accessed self.train_targets and self.train_inputs unconditionally, which failed for ApproximateGPyTorchModel because its training data lives on its model attribute (e.g. model.model.train_targets).

Changes:

• Extracted two new overridable methods on GPyTorchModel: _untransform_targets() (undo outcome transform and return raw targets) and _retransform_and_set_targets(). Note: None typing effectively replaces hasattr checks for train_targets
• Used these methods in load_state_dict
• Overrode both methods on ApproximateGPyTorchModel to read from / write to self.model.train_targets and self.model.train_inputs instead of self.

Differential Revision: D98021112

[meta-codesync]

@esantorella has exported this pull request. If you are a Meta employee, you can view the originating Diff in D98021112.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 99.98%. Comparing base (42facfa) to head (cfe0514).

Additional details and impacted files

@@           Coverage Diff           @@
##             main    #3251   +/-   ##
=======================================
  Coverage   99.98%   99.98%           
=======================================
  Files         220      220           
  Lines       21871    21890   +19     
=======================================
+ Hits        21867    21886   +19     
  Misses          4        4           

☔ 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.
• 📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.