Implementing USM memory management in dpCtrl

[diptorupd]

PyDPPL should expose sycl USM allocators to Python.

What needs to be done?

1. Add a C-API wrapper into pydppl/backends for the functions that we need from cl::sycl::usm. The C-API will be similar to what we currently have for cl::sycl::queue (ref. https://github.com/IntelPython/pydppl/blob/master/backends/include/dppl_sycl_queue_interface.h).

2. We need this so that Sycl headers do not have to be directly included in the Cython extension module and we can use MSVC/GCC to build the extension and link to our helper (libDPPLSyclInteface).

3. Cython extension:

• We need three classes: MemoryUSMShared, MemoryUSMHost, MemoryUSMDevice. Each class corresponds to the three USM memory allocator.
• MemoryUSMDevice should not expose __getbuffer__ as Python objects cannot directly use device memory.
• The __cinit__ of the Memory classes perhaps should provide a way to specify the context in which memory is to be allocated. The other option will be to use dppl.get_current_queue(). Using the current queue always can lead to programmer errors where he/she tries to use a memory allocated on a given context on a device that is not in that context.
• MemoryUSMxxx should provide a factory function to copy data between incompatible device.

[diptorupd]

@PokhodenkoSA Look at #30. It introduces changes in how a Sycl queue is stored in dppl. I have added a new class SyclQueue that replaces the previous Py_Capsule. The changes were suggested by @oleksandr-pavlyk to make it easier for us to get the sycl::context from the queue. You can now use SyclQueue::get_sycl_context().get_context_ptr() to get an opaque reference to the context pointer of the queue.

Please use the new SyclQueue class in your memory module. We most probably need a sycl_core.pxd to provide you with the Cython header exposing SyclQueue and SyclContext. I will absolutely not mind if you create that pxd file before I get to it 😉.

[diptorupd]

Can we expose the USM fine-grained control functions mem_advise and prefetch in the C API? We might want to use them in Numba.

[PokhodenkoSA]

This methods will belong to SyclQueue and will get MemoryUSM classes. Is it correct?