USM support

[igchor]

UMF should support managing USM (unified shared memory) by:

• exposing memory pools capable of handling memory that might not be accessible on the host:
  • [x] disjoint_pool
  • [x] jemalloc_pool
    • note: there is a known issue with jemalloc: if compiled with default flags it cannot be dlopened (neither can a library that links with jemalloc) which makes it problematic to use in Unified Runtime which manages adapters using dlopen
• exposing memory providers capable of allocating memory on a device:
  • [x] L0 provider (currently implemented in Unified Runtime)
  • [ ] CUDA provider (currently implemented in Unified Runtime)
  • [ ] Unified Runtime (?)

It should be possible to build L0 and CUDA providers as static libraries so that they can be used by Unified Runtime.

We should also consider:

• adding memory copy / memset operations for providers (to implement calloc/realloc on device-only memory)
• async memory management (transfers and/or allocation)

[vinser52]

Another requirement from MPI/oneCCL. They do not have hard dependency to L0 today the same MPI/oneCCL binaries should work on the system without L0. They resolve dependency to L0 at runtime using dlopen/dlsym to look for libze_loader.so. We should consider this case when developing L0 memory provider.