Rework occupancy, re-enable grid-stride broadcast

[maleadt]

The occupancy API (launch_heuristic, launch_configuration) needed an update because oneAPI's occupancy API wants to know about the total width of the launch, which wasn't passed to launch_configuraton. I've taken that opportunity to rename total_threads to elements.

I also came across the disabled support for grid-stride broadcast, where each thread performs a couple of iterations in a loop instead of just launching more blocks. I didn't notice performance improvements when initially adding this, but doing some more careful benchmarking it turns out to resolve the performance issue of launching as many threads as the CUDA occupancy API suggests. My hypothesis is that it's expensive to launch many large blocks, so we capped the block size to 256, whereas with grid-stride loops it's instead possible to keep the block large but launch fewer of them.

The above, in benchmark results of broadcasting sin over a 1024x1024 Float32 array:

• 4096 blocks of 256 threads (the current situation): 25 us
• 1024 blocks of 1024 threads (what the occupancy API suggests): 30 us
• 48 blocks (as suggested by the occupancy API) of 256 threads (i.e. on top of the current situation), each doing 86 elements: 25 us, so no performance improvement, and why I didn't enable this functionality
• 48 blocks of 1024 threads (as suggested by the occupancy API), each doing 22 elements: 25us

Even though the combination of a grid-stride loop with larger blocks doesn't yield a performance improvement, it does stay closer to the NVIDIA-recommended configuration, so I'm inclined to use that configuration instead.