Reduce the number of kernels to compile in merge-sort

[dmitriy-sobolev]

This PR implements a TODO:

    // TODO: split the submitter into multiple ones to avoid extra compilation of kernels:
    // - _LeafSortKernel and _CopyBackKernel do not need _IndexT

It results in reduced kernel compilation times. The table below shows a JIT compilation speed up for different devices:

	Named Kernels: JIT speedup, times	Unnamed Kernels: JIT speedup, times
Intel® Data Center GPU Max Series (zeModuleCreate calls)	1.18	1.61
Intel(R) Xeon(R) Platinum 8480+ (clBuildProgram calls)	1.26	1.85

There is no observable impact on the execution time besides kernel compilation.

I used the following program as a benchmark:

#include <sycl/sycl.hpp>

#include <oneapi/dpl/algorithm>
#include <oneapi/dpl/execution>

#include <iostream>
#include <algorithm>

int main()
{
    using T = int;
    const int n = 50'000;
    sycl::queue q;
    T* data = sycl::malloc_shared<T>(n, q);
    std::generate(data, data + n, [n = 0]() mutable { ++n; return n % 1000; });
    // Named Kernels
    oneapi::dpl::sort(oneapi::dpl::execution::make_device_policy<class sort>(q), data, data + n, [](const T& a, const T& b) { return a < b; });
    // Unnamed Kernels
    // oneapi::dpl::sort(oneapi::dpl::execution::make_device_policy(q), data, data + n, [](const T& a, const T& b) { return a < b; });

    std::cout << std::is_sorted(data, data + n) << std::endl;
    sycl::free(data, q);
}