Add standalone puppi, some fixes to the bucketing in HEPT, a few addition to the standalone script

[kawaho]

1. PUPPI

• Added puppi.py, which is a standalone puppi script that runs on numpy arrays with no CMSSW/MLPF dependence
• eval.py now runs puppi and save the results to parquet file for analysis

2. HEPT

• Increment hashed index along phi first, then eta so that tracks/clusters in opposite ends of eta won't be grouped together
• Allow hash index to go across pi to -pi
• Apply random shift to phi so that the buckets do not necessarily start at phi=-pi
• Save the bucket index and coordinates to parquet files to understand how regions are grouped in HEPT

4. General

• Add evaluation only mode to just run evaluation only on saved model
• Added a neutral PU FFN
• Save the attention matrix to npz files for the ease of changing the plotting style

[jpata]

Did you plan to commit the change of input features? That requires a change on the CMSSW side as well.
I think in general it makes sense to test new models from new features separately to avoid entangling effects.

[kawaho]

I agree. I can do that in a separate PR. I should be finished with this one within today.