ThresholdChecker updates

ThresholdChecker is updated to compute improvement according to the last improved step and not to the best-received metric.
ThresholdChecker(0.5), scores 0.2->0.4->0.8
Previously ---> no improvement, no improvement
Now           ---> no improvement, IMPROVED

Dense & StatsPrint

• Dense custom layer to support applying norm (configurable to before or after activation)
• StatsPrint callback to support printing best-confusion-matrix when at least one of the metrics is of type MetricConfusionMatrixBaseSome
• minor cosmetic changes

Activation in TransformerEncoderStack

• TransformerEncoderStack to support activation as input
• PositionalEncoding to support more than 3 dimensions input

Minor fixes - 2023 happy new year

• Added assert to Attention class (from extensions) when mask is used
• Fixed confusion matrix cpu/gpu device error
• Better handling on callbacks where apply_on_states=None (apply on all states)
• Updated Pipfile

AbsoluteThresholdChecker & RelativeThresholdChecker

• Added AbsoluteThresholdChecker & RelativeThresholdChecker classes
• ThresholdCheckers can now be used in CallbackMonitor to better define metric tracking

Validation fix

Bug fix in case validation samples are empty

Validation fix

Bug fix in case validation samples are empty

Verbosity

Bug fix in verbosity level 2 in train

Verbosity change in torch_utils

verbosity default to false in some functions

Fix to PositionalEncoding to be batch first

As well as taking the updated PositionalEncoding implementation from torch official tutorial