Set clamped vs unclamped FAPE for each sample in batch independently

[ar-nowaczynski]

Previously, whether to use clamped vs unclamped FAPE backbone loss was decided once for all samples in the mini-batch.

Now, clamped vs unclamped FAPE backbone loss is decided for each sample in the mini-batch independently. This brings substantial improvements to training convergence and stability.

[gahdritz]

Thanks Arkadiusz!