Bugfix/from model plot

[cregouby]

Fix #148 and fix #149.

• Remove the unused $loss level in the list of collected metrics along training. This simplifies the data wrangling for ggplot::autoplot() as well.

The following training sequence now produce the expected plot

library(tabnet)
data("ames", package = "modeldata")
fit_no_checkpoint <- tabnet_fit(Sale_Price ~., data = ames, epochs = 14, valid_split = 0.2, checkpoint_epoch = 15, batch_size = 64)
fit_with_checkpoint <- tabnet_fit(Sale_Price ~., data = ames, tabnet_model = fit_no_checkpoint,epochs = 6, checkpoint_epoch = 1)
autoplot(fit_with_checkpoint)