feat(eyeTracking): improve batch_predict visualization payload (heatmap + points)

[harsshks]

Summary

This PR improves the /api/session/batch_predict response to support richer eye-tracking visualizations (heatmaps and prediction/scanpath points) while preserving backward compatibility.

What changed

• Introduced a lightweight heatmap generator (app/services/heatmap.py) that:
  • builds a 2D gaze density map from prediction points
  • applies optional Gaussian smoothing
  • normalizes values to the 0–1 range for frontend rendering
• Restructured predict_new_data_simple to return a unified, self-descriptive payload containing:
  • per-prediction data (x, y, x_norm, y_norm, timestamp, confidence)
  • screen dimensions
  • a heatmap placeholder
• Updated batch_predict route to attach heatmap data safely (wrapped in error handling so predictions still return if heatmap generation fails)

Resulting response shape

{
  "screen": { "width": W, "height": H },
  "predictions": [
    { "x": 678.6, "y": 383.7, "x_norm": 0.49, "y_norm": 0.50, "timestamp": 123456, "confidence": 0.87 }
  ],
  "normalized": true,
  "heatmap": [[...]]
}