feat: linarize most bached simulation kernels by switching to batch-interleaved formats

[sebcrozet]

This PR Modifies the way internal objects in the rigid-body simulation (contacts, constraints, collision pairs, etc.) are stored in gpu memory. Before, we preallocated a range of contiguous elements per-batch. For example elements 0..n are for batch 0, elements n + 1..2n for batch 1, etc. While the mental model for this is simple, this makes the indexing complicated in scenarios where different batches might contain different numbers of elements (meaning that, for example, elements k..n are empty).

This PR switches these to batch-interleaved format. This is essentially the transpose of the above. It means that elements 0, nb, 2nb, 3nb, ... are for batch 0 (where nb is the number of batches), elements 1, nb + 1, 2nb + 1, 3nb + 1, ... are for batch 1, etc. So the stride between two elements belonging to the same batch is equal to the number of batches. The consequence is that all empty entries are pushed to the end of the buffers, making indexing much simpler. It also means that several kernels can just iterate through the elements linearly, without worrying about the concept of batches at all.