Training for hit cluster classification FlagBkgHits

[NamithaChitrazee]

MLP and BDT based training for hit cluster classification.
Input dataset for the training can be created by running ArtAnalysis/fcl/BkgDiag.fcl on a digitized data sample.
CreateInference.C converts the .h5 output of the training to a .hxx file that can be used in Mu2e Offline.

[oksuzian]

Took a look through this PR. A few things worth addressing before merging:

Bugs

• Input file list doesn't match what was actually trained on. KKTrainBkgFiles.txt contains /TrainBkg/RootFiles/nts.owner.BD.version.sequence.root, but the committed notebook outputs show the real paths are under /Users/namithachithirasree/Documents/LBL_projects/KKTrain/RootFiles/.... Anyone re-running this will fail immediately.
• Pre-selection stats only reflect the last file processed. The cell that computes kQpresel = kQ[kQ==-1] / kQpass = kQ[kQ>-1] uses the loop-local kQ (the last file, pbar). It should use the concatenated kQ_all.
• Hardcoded absolute user paths leaked via notebook outputs (/Users/namithachithirasree/...). Worth stripping outputs before commit.

Redundancy

• output_model.save("model/TrainBkgDiag.h5", ...) and model_ce.save("TrainBkgDiag.h5") save two copies; the cwd one looks unused.
• fpr_ce, tpr_ce, th_ce = roc_curve(...) and auc_ce = roc_auc_score(...) are computed twice in back-to-back cells.

Code quality

• Eight for i in range(arr.shape[0]): list.append(arr[i][k]) blocks could each be arr[:, k].
• Train/test/valid ends up as 50/25/25 (two successive test_size=0.5 splits). Was that intentional? 70/15/15 is more typical.
• Several unused imports: os, datetime, Sequential, Dropout, Activation, ReLU, SGD, duplicate Adam.
• output_model (batch=1) duplicates model_ce (batch=bsize) — any architecture change has to be made in both places to keep SOFIE export working.

Missing

• No TrainBkg/README.md. Both TrkQual/ and the in-flight TrkPID/ PR have one — would be good to document datasets used, the feature list, SOFIE/ROOT version, and rerun instructions.
• CreateInference.C has an inline "In 6.30…" note but there's no README pointer to required ROOT version.