Use std::size_t for requested scratch allocations in GPU backends

[masterleinad]

In #4532, it turned out that some test cases needed the allocation size for scratch space to be larger than ExecutionSpace::size_type. This pull request tries to make the change from ExecutionSpace::size_type to std::size_t more uniform across CUDA, HIP and SYCL.

[cz4rs]

see below (this comment cannot be deleted)

[cz4rs]

scratch_unified definition was not updated:

/var/jenkins/workspace/Kokkos/core/src/Cuda/Kokkos_Cuda_Instance.cpp:623:32: error: out-of-line definition of 'scratch_unified' does not match any declaration in 'Kokkos::Impl::CudaInternal' [clang-diagnostic-error]

Cuda::size_type *CudaInternal::scratch_unified(

                               ^

Also these look to be mismatched:

kokkos/core/src/Cuda/Kokkos_Cuda_Instance.hpp

Lines 63 to 68 in 83acf58

	CudaSpace::size_type* cuda_internal_scratch_flags(
	const Cuda&, const CudaSpace::size_type size);
	CudaSpace::size_type* cuda_internal_scratch_space(
	const Cuda&, const CudaSpace::size_type size);
	CudaSpace::size_type* cuda_internal_scratch_unified(
	const Cuda&, const CudaSpace::size_type size);

Otherwise looks good to me.