normalize loss, reparametrize network

[jpata]

• normalize loss function for classification by the number of not padded elements
• normalized loss function for regression by the number of true target particles and with the stddev of the regression targets for more stable loss values
• reparametrize attention network to have num_heads, head_dim
• produced new version 1.7.1 of cms_pf_multi_particle_gun with more stats
• remove unneeded pad_power_of_two option (seems like FlashAttention does that internally), not sure why I thought it was needed
• make regression output type configurable
• disable charge prediction for now (so far we didn't really study its performance)
• enable setting only certain layers as trainable
• set minimum lr to 1e-5 for cosinedecay
• add new standalone notebook for quick studies
• remove TF workflows from pipeline