Stepping through the rows by batch_size leaves a short final batch, and a
final batch of one row is worse than short. nre()'s atomic loss has no
contrast to score a single simulation against, so it returns a constant that
torch did not build and backward() errors with "element 0 of tensors does
not require grad"; every other estimator takes a gradient step from one
sample. Those rows join the previous batch instead. It costs one batch of
batch_size + 1 once an epoch and removes a failure that depends on nothing
but n_simulations modulo batch_size.