Atlernative allocation strategies

[maleadt]

CLArray currently uses coarse-grained SVM for device allocations, so that it's possible to get a hold of device pointers (instead of opaque handles) for Julia ABI reasons. Some back-ends may not support this, so it would be interesting to have alternative allocation strategies:

• using fine-grained SVM: although overkill, it may enable additional platforms
• using USM: https://registry.khronos.org/OpenCL/extensions/intel/cl_intel_unified_shared_memory.html (also supported by pocl)

I personally don't have a need for this, so won't be working on this, but I'd be happy to take patches that extend compatibility of CLArray.