use SYCL queue.memcpy instead of copy kernel

[jianyizh]

depend on jira GSD-3944. It's expected that d2d copy through copy engine and kernel should be the same performance in level zero 2, but I still see gaps.

Note memcpyAsync is only called when memcpy_eligible, means no cast, and all contiguous, so it's just copy without other operations. Previous implementation in copy_kernel calls two implementation by loops kernel: gpu_kernel(iter, CopyScalarFunc<scalar_t>()); and float8_copy_kernel_xpu, since we won't do cast, they both equivalent to queue.memcpy

[Copilot]

Pull Request Overview

This PR updates the XPU copy implementation to use SYCL queue.memcpy instead of copy kernels for better performance. The change addresses performance gaps between device-to-device copy through copy engine and kernel implementation in Level Zero 2.

• Simplifies the memcpyAsync function by removing the device comparison branch and always using q.copy()
• Removes outdated performance-related comments that are no longer relevant
• Maintains the same logical flow while leveraging improved SYCL queue.memcpy performance

---

_{Tip: Customize your code reviews with copilot-instructions.md. Create the file or learn how to get started.}

[Copilot]

The function now always calls q.copy() regardless of whether devices are the same or different. For same-device copies, this may not be the intended behavior since the original code used copy_kernel(iter) for same-device operations. Consider verifying that q.copy() handles same-device copies correctly or add a comment explaining why this unified approach is safe.

Suggested change

	if (dst_device != src_device) {
	TORCH_INTERNAL_ASSERT(p2p_enabled == true);
	auto dst = (char*)iter.data_ptr(0);
	auto src = (char*)iter.data_ptr(1);
	size_t size = iter.numel() * iter.element_size(0);
	auto q = copy_stream.queue();
	q.copy(src, dst, size);
	}
	auto dst = (char*)iter.data_ptr(0);
	auto src = (char*)iter.data_ptr(1);
	size_t size = iter.numel() * iter.element_size(0);
	auto q = copy_stream.queue();
	q.copy(src, dst, size);
	auto dst = (char*)iter.data_ptr(0);
	auto src = (char*)iter.data_ptr(1);
	size_t size = iter.numel() * iter.element_size(0);
	auto q = copy_stream.queue();
	if (dst_device == src_device) {
	// Use the optimized kernel for same-device copies
	copy_kernel(iter);
	} else {
	TORCH_INTERNAL_ASSERT(p2p_enabled == true);
	q.copy(src, dst, size);
	}