Fine-tune the Masked Particle Modeling foundation model for tau lepton reconstruction

[Laurits7]

See here for MPM: https://arxiv.org/pdf/2401.13537

Following the example of OmniParT: https://github.com/HEP-KBFI/ml-tau-en-reg/blob/main/enreg/tools/models/OmniParT.py
define a model where the pretrained MPM encodings are used as an input to the binary classification, multiclassification and regression heads.

Compare a pretrained model with fixed weights, to a model with fine-tuned weights, to a model trained from scratch.
Compare physics metrics (cls, reg resolution) as a function of the training statistics.