sycl

This patch adds template argument to the sycl::buffer_allocator class (according to SYCL 2020 spec, sycl::buffer_allocator is templated on the data type.) The new sycl::buffer_allocator is defined as a class to let C++17's class template argument deduction feature work. So, the old code that uses `sycl::buffer_al

Bug summary
There is evidence that sub_group::get_group_id() does not return the same value as threadIdx.x / warpSize (assuming 1D kernel), as expected on CUDA. We should check the implementation of this function. Our implementation of this function performs bit manipulation magic, presumably the optimization went to far...

To Reproduce
Compare sub_group{}.get_group_id() or `sub

The problem is that the OpenCL types in https://github.com/triSYCL/triSYCL/blob/master/include/triSYCL/opencl_types.hpp are defined on the host according to the x86-64 Linux ABI which depends on the CPU & OS instead of using the description from https://www.khronos.org/registry/OpenCL/specs/2.2/html/OpenCL_C.html#built-in-scalar-data-types

Note that the system-wide cl_size_t has been removed

The cl_ types are not needed in general, however they should still be kept in the OpenCL C interop sample code.

Things like XILINX_XRT which are set by the user and others like XILINX_DEVICE_ONLY that are set by the compiler but affect the code users compile may be useful to add.

DEBUG variable should be removed from numba_dppy.compiler.
It will require replace callers code.
It is for separate PR.

Originally posted by @PokhodenkoSA in IntelPython/numba-dppy#354 (comment)

Some of the gt-blas calls require temporary space, particularly the SYCL backend. Having a managed storage helper class would be useful for implementing this, in a nice RIAA way.