feat: Retrain Attention ResUNet segmentation model and resolve dimension mismatch

[Nicks-19]

Overview

This Pull Request updates the 162-layer Attention ResUNet brain tumor segmentation model. It uses a dataset and it fixes some problems, with the Flask backend. This fixes issue #74.

Key Modifications

1. Model Architecture & Compatibility Updates (app.py)

The code now looks for ResUNet-weights.keras first when it loads the segmentation model. If it does not find this it will use weights_seg.hdf5 instead.
This fixes a problem where the wrong number of layers were loaded. This caused an error that said ValueError: Layer count mismatch....

2. Loss & Metrics Precision Fixes (utilities.py)

The code now makes sure that y_true and y_pred are the type when it calculates the loss and metrics. This prevents a TypeError from happening when the code is run.
This is because Keras 3 and TensorFlow 2.x need float32 but the code was giving them int64.

3. Standalone Training Pipeline (train_segmentation.py. ResUNet-model.json`)

There is a Python script that can train the ResUNet model on its own. It builds the model trains it and saves the weights.
The script can train both the 90-layer simple ResUNet and the 162-layer Attention ResUNet.
The ResUNet-model.json file has been updated to make sure everything is consistent.

4. Model Ignored in Git (.gitignore)

The .gitignore file now tells Git to ignore *.keras and *.h5 files. These are the weight files that we do not want to commit to GitHub.

---

Verification and Metrics

1. Flask App Initialization

Verified that the backend starts up and loads all ensemble and segmentation models successfully without any dimensions mismatch:

Loading primary classification model...
✓ Primary classification model loaded successfully
Loading secondary classification model...
✓ Secondary classification model loaded with JSON architecture
Loading primary segmentation model (ResUNet-MRI)...
✓ Primary segmentation model loaded successfully
Loading secondary segmentation model...
✓ Secondary segmentation model loaded successfully
    Model Summary:
   - Classification models loaded: 2
   - Segmentation models loaded: 2
   - TTA enabled: True
   - Ensemble enabled: True
All models loaded successfully: True

[Brijeshthummar02]

@Nicks-19 add issue no in pr description