Remove loading pretrained weights during inference pipeline

[gitttt-1234]

This PR fixes checkpoint loading in the inference pipeline: we shouldn't reapply pretrained_backbone_weights when constructing the Lightning module for inference. Only the trained checkpoint state is loaded, which avoids accelerator/device mapping errors (e.g., trained on GPU, inferred on CPU) that arose from tying pretrained weight loading to trainer_accelerator. This PR makes inference device-agnostic and ensures only the trained weights are loaded.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 95.39%. Comparing base (ff91433) to head (5011a3d).
⚠️ Report is 20 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #305      +/-   ##
==========================================
+ Coverage   95.28%   95.39%   +0.10%     
==========================================
  Files          49       49              
  Lines        6765     6990     +225     
==========================================
+ Hits         6446     6668     +222     
- Misses        319      322       +3     

☔ 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.