Feature/gaussian point model

[sridhs21]

Reworks the X-point detector to train on Gaussian heatmap targets and evaluates it at the point level instead of per pixel. The hyperparameter-tuning work this branch was originally stacked on has since merged in #21, so this PR now contains only the point-level and Gaussian changes.

Gaussian heatmap model (XPointMLTest.py)

• --targetType gaussian renders each ground-truth X-point as a 2D Gaussian peak (--gaussianSigma, default 3) instead of the dilated 9×9 binary mask; binary is still the default
• --lossFunction heatmap_focal: CornerNet/CenterNet-style penalty-reduced focal loss, with --heatmapAlpha / --heatmapBeta

Point-level evaluation

• point_metrics.py: extract_peaks (vectorized connected-component NMS, one peak per region above the threshold) and match_points (greedy one-to-one matching within a radius, highest confidence first), plus aggregation over frames
• run_hessian_and_build_cache.py: also writes a ground-truth {N}_xpts.csv (row, col, class) for every cached frame
• backfill_xpts_csv.py: writes those CSVs for caches built earlier, with --output-dir for caches that aren't writable
• predict_points.py: runs a checkpoint over cached frames and writes per-frame predicted {N}_xpts.csv files with confidences
• score_point_predictions.py: matches predictions to ground truth and reports F1 / precision / recall / TP / FP / FN per dataset

Tuning

• optuna_tuner_gaussian.py: Optuna tuner for the Gaussian model that maximizes point-level F1 on the validation frames rather than minimizing validation loss
• The peak-extraction threshold is swept inside each trial's F1 evaluation (--threshold-grid) instead of being fixed at 0.3, so every configuration is scored at its own best operating point; the winning threshold is recorded per trial

Data and configuration

• Second, independent PKPM run (PKPMv2) supported in cache building, prediction, and scoring
• Dataset paths in predict_points.py and score_point_predictions.py are built from RC_EXTRACT_DIR / RC_CACHE_BASE / RC_PKPM_V0_ROOT (the same convention as run_hessian_and_build_cache.py), with the shared DeltaAI locations as defaults, so launchers can point them at another environment without code changes
• test_xpoint_transfer.py: THRESHOLD environment variable for the pixel-level transfer evaluation

README

• Documents the new target and loss options, the Gaussian tuner and threshold sweep, PKPMv2 and the CSV outputs, the point-level evaluation pipeline, and the path environment variables

[cwsmith]

Thank you. With the change to the model, are there changes needed in the top level README (e.g., new command line args, etc.)?