Update memory_index once per batch, not per tensor

[AlexD15216]

The write pointer was being incremented inside the for-loop that iterates
over every tensor in the batch. When more than one tensor is present (common situation in RL), this
advanced the pointer multiple times for the same batch, leaving gaps and
eventually skipping slots in the ring buffer.

Because all tensors share the same first dimension (batch size), the
pointer should move once after the entire batch has been copied.
The increment logic is now placed after the loop so the index grows
by num_samples exactly once.