draft of attention based injection gate

Closes #50 (closed)

It was found that the original model for the injection prediction tends to unlearn predicting the timing of injection background spikes based on BEAMGATESTATUS PV of prev 9seconds.

The plot below illustrates the issue. the model was training on the first half of June 2021 and used to predict on 28th of June 2021. The prediction of injection timing is pretty poor (while looking good during the training time window).

The new model (using an attention mechanism to learn attending to the correctly time lagged Gate Status PV) masters the task much better based on the same training.

(@natochii )