2. Are rigid bodies collisions accelerated throught OpenCL? Or it's just cloth simulation?
No. This is just very early work that we wanted to get into the tree quickly for various reasons and to see what changes we'd need to make within the bullet code to cleanly integrate accelerated computation engines. The DirectCompute version has progressed further with more optimisation of data access patterns to play with how to perform the solutions efficiently on a multithreaded wide SIMD machine. The optimisations in the DC version can be carried across to CL with minimal effort but I've not had time to test it yet.
I do have a GPU-emulation version of one-way capsule collision now that integrates into the Bullet broadphase. This will map across to actual GPU compute kernels very simply but again, just like the GPU I'm multitasking on this. Expanding it to other colliders is not a serious complication but doing two-way collisions efficiently will still need some more thought.
I can't comment on issues with the particle sim or Apple problems. I have seen an early version of this run on a mac, though, Erwin and my colleague had it running some time ago. It's possible that something changed or the changes to the build system broke something.