Only compute step loss on non-same-day reviews

[joshdavham]

Previously, the optimizer would only compute step loss on review-state cards. For example:

if card.state == State.Review:
    step_loss = loss_fn(y_pred_retrievability, y_retrievability)

However, it makes more sense to more generally compute step loss on non-same-day (long term) reviews rather than depending on the specific learning state of the card. Technically, in the past, a user could review a review-state card multiple times in one day and this fix simply accounts for that case.

Now, step loss is computed on all non-same-day reviews, not just review-state cards.

if card.last_review and (review_datetime - card.last_review).days > 0:
    step_loss = loss_fn(y_pred_retrievability, y_retrievability)

I also adjusted one of the unit tests as the optimal computed parameters are now slightly different.

Let me know if you have any questions 👍

[L-M-Sherlock]

LGTM

[L-M-Sherlock]

For the calculation of elapsed days, you can refer to:

• Fix/wrong calc elapsed days ts-fsrs#144