ARM compatibility

[John-194]

ARM CPUs need strict memory orders. While x86 has a stronger memory model, it should not be relied upon.

Specifically, in this case, local_bot and done cannot be relaxed for ARM; some other values could possibly be relaxed (like a simple finished_flag), but to be on the safe side, I changed other ones too from relaxed to release/acquire.
Performance-wise, memory_order_acquire and memory_order_release should be unnoticeable on x86, while possibly adding some safety.

[skbill84]

Hi @John-194 ,
Thank you for initiating this fix. We have a question - under what scenario will the code run into an ARM related crash? Could you share the recipe to reproduce it?

[John-194]

Hi @John-194 , Thank you for initiating this fix. We have a question - under what scenario will the code run into an ARM related crash? Could you share the recipe to reproduce it?

I tested on a Jetson. I ran DBSCAN on a point cloud on repeat without any thread restrictions. It might not crash right away, but after a few iterations, the script seg faults or freezes. There is also a related open issue.