add vectorization path on maxpool forward channel last

[jianyizh]

Part 1 of #1861
tested on shapes from alexnet training
on BMG, 831719 Scoreboard stalls decrease to 497,098. instruction fetch and distance stall also get better.

shape	device	before opt	after opt
[4096, 64, 55, 55]	pvc	8.02 ms	5.44 ms
[4096, 64, 55, 55]	bmg	12.45 ms	8.89 ms
[4096, 192, 27, 27]	pvc	5.72 ms	3.85 ms
[4096, 192, 27, 27]	bmg	9.00 ms	5.06 ms
[4096, 256, 13, 13]	pvc	1.68 ms	1.12 ms
[4096, 256, 13, 13]	bmg	2.83 ms	1.35 ms

[Copilot]

Pull Request Overview

This PR adds a vectorized code path for the max pooling forward operation when using channel-last memory layout, providing significant performance improvements on Intel GPU architectures. The optimization uses vectorized memory operations and SYCL kernels to improve throughput.

Key changes:

• Introduces a new vectorized kernel MaxPool2dChannelLastVec that processes multiple channels simultaneously
• Adds automatic vector size selection (8, 4, 2, or 1) based on data alignment and hardware capabilities
• Implements dynamic work group sizing based on hardware thread availability

[Copilot]

[nitpick] The complex offset calculation spans multiple lines and lacks spacing around operators. Consider breaking this into intermediate variables or adding consistent spacing (e.g., 'inputSizeH_ * inputSizeW_ * numPlane_').

Suggested change

	load_offset = batch * inputSizeH_*inputSizeW_*numPlane_ / vec_size + plane +
	h * inputSizeW_ * numPlane_ / vec_size + w * numPlane_ / vec_size;
	int64_t batch_offset = batch * inputSizeH_ * inputSizeW_ * numPlane_ / vec_size;
	int64_t plane_offset = plane;
	int64_t height_offset = h * inputSizeW_ * numPlane_ / vec_size;
	int64_t width_offset = w * numPlane_ / vec_size;
	load_offset = batch_offset + plane_offset + height_offset + width_offset;

[Copilot]

Using the private '_isnan' function instead of the standard 'std::isnan'. Consider using 'std::isnan' for better portability and standards compliance.

[Copilot]

C-style cast to 'char*' should be replaced with a safer C++ cast like 'reinterpret_cast<char*>(input)' for better type safety and clarity.

Suggested change

	std::min(8, memory::can_vectorize_up_to<scalar_t>((char*)input));
	std::min(8, memory::can_vectorize_up_to<scalar_t>(reinterpret_cast<char*>(input)));