Autoencoder, ATM and global optimizer

See merge request !24

Simulating parallel execution

Closes #1

See merge request !23

Fix lr schedule

See merge request !22

Adjusted FLOP estimation to account for different parts of the model used in forward and backward pass

This is a quick and dirty fix to make it easier to run experiments.
A proper implementation should relocate the parallelism into the
corresponding dispatcher classes.

Closes #1

Simulating parallel execution by running everything sequentially while measuring the time and computing the expected parallel time

Fixed AE weights loading issue

See merge request !21

Removing points from experiment group names to fix model saving and loading

See merge request !20

PSL performance

See merge request !19

Merge in main

See merge request !18

# Conflicts:
#   README.md

Fix PSL

See merge request !16

Added CIFAR-10 dataset

See merge request !17