Adding LZ extension zexDriverImport/Export of ExternalPointer to hipHostRegister

[colleeneb]

This continues to address #1260 .

This PR adds importHostMemory and exportHostMemory to the hipHostRegister and hipHostUnregister calls. For LZ, this then calls a LZ extension (https://github.com/intel/compute-runtime/blob/master/programmers-guide/SYSTEM_MEMORY_ALLOCATIONS.md) that improves the speed of copies. The test tests/benchmarks/hostRegisterMemcpyOverhead.hip shows a speedup for using hipHostRegister for a host memory before memcopying it many times compared to not using hipHostRegister on PVC on Aurora at least:

1000 hipMemcpy H2D (4 MB) WITH    hipHostRegister: avg=0.117369 ms  min=0.11494 ms  max=0.964309 ms
1000 hipMemcpy H2D (4 MB) WITHOUT hipHostRegister: avg=0.254521 ms  min=0.242622 ms  max=0.345852 ms
overhead factor (avg): 0.461138x

(The first call is slower with hiphostregister, but subsequent calls are ~2x faster than without.)

If the extension is not found, it will silently continue without.

Once we get a newer Level Zero release that contains version 1.14.33, we can switch to using ze_external_memmap_sysmem_ext_desc_t where we will be able to directly map malloc'd memory to USM host memory, but for now this should help a bit.

[pvelesko]

/run-aurora-ci