Feature/159 psf output from trained model

[jeipollack]

This PR adds inference capabilities to the WaveDiff / WF-PSF framework, allowing the user to generate broadband (polychromatic) PSFs from a trained model, given a list of positions and SEDs.

It had a refactor PR merged into it in order to carry on with the development which is why it contains >100 commits (unfortunately and not a practice that will be repeated).

Closes #159 (Validation runs described there)

Key Changes

• Introduced psf_inference.py and associated test_inference packages for PSF inference workflows
• Refactored TFPhysicalPolychromatic and related modules to separate training vs. inference behaviour
• Added PSFInferenceEngine class to centralise training, simulation, metrics, and inference logic
• Moved model weights loading to psf_model_loader.py
• Centralised PSF data extraction in data_handler
• Updated configuration handling for inference (inference_config.yaml)
• Updated documentation (small) to include inference package
• Core Refactoring from PR 165 refactor tfphysicalpolychromatic to cleanly separate training vs inference behaviour #166:
  • New run_type attribute: Added to data_handler supporting training, simulation, metrics, and inference modes
  • ZernikeInputsFactory class: New factory class that builds ZernikeInputs instances based on the specified run type
  • Enhanced ZernikeInputs data class: Now intelligently assembles Zernike contribution lists based on run_type and available data (including priors and source positions for CCD misalignment corrections)
  • Hybrid loading pattern: Refactored to use eager loading in constructors combined with lazy-loading via property decorators, reducing constructor overhead and improving modularity
  • Code organization: Improved module structure including new tf_utils.py module in the psf_models sub-package
  • Bug fix: Fixed incorrect argument name in DataHandler class that prevented training dataset images from being converted to TensorFlow data type as expected
  • Helper Methods: Added several helper methods to assist with preparing datasets and ensure data type conversion is performed as expected
  • Deprecations: Removed get_obs_positions as it was no longer needed
• Bug fix: Fix logger formatting for relative RMSE metrics in metrics.py (values were not being displayed)
• Changelog Fragment added describing updates

Testing

• CI tests pass successfully
• Unit tests cover:
  • Loading trained models and other aspects of inference setup
  • Generating PSFs for multiple positions/SEDs and error handling
  • Consistency of outputs and tensor handling
  • Clearing cached values
• Verified backward compatibility with existing develop code
• Confirmed inferred PSFs from test datasets match expected results with high accuracy

[jeipollack]

it would be very ideal to squash and merge this commit in order to simplify rebasing the other branches.