add vectorization path on maxpool backward channel last

[chunhuanMeng]

Part 2 of #1861
on PVC, 101,628 Scoreboard stalls decrease to 75,976. Significantly fewer instruction fetch and distance stalls, enabling higher effective bandwidth to HBM.

shape	device	before opt	after opt
[4096, 64, 27, 27]	pvc	27.10ms	12.70 ms
[4096, 192, 13, 13]	pvc	17.97ms	8.51 ms
[4096, 256, 6, 6]	pvc	5.10 ms	2.47 ms

[Copilot]

Pull Request Overview

This pull request adds a vectorization path for maxpool backward operations in channel-last memory layout to improve performance. The change introduces a new templated kernel implementation that processes multiple elements simultaneously using vector operations.

• Refactors existing backward kernel to accumulate gradients locally before writing
• Adds new vectorized kernel implementation for channel-last memory layout
• Includes vectorization logic (currently commented out) with macro for launching vectorized kernels

[Copilot]

The vec_size parameter should be 4, not 1, for the case 4 branch. This appears to be a copy-paste error that would prevent proper vectorization when vec_size is 4.

Suggested change

	// 1,
	// 4,

[Copilot]

The cast static_cast<scalar_t>(grad_vec[i]) is redundant since grad_vec[i] is already of type scalar_t. This should be simplified to grad_vec[i] += static_cast<scalar_t>(gout_val_vec[i]);

Suggested change

	grad_vec[i] = static_cast<scalar_t>(grad_vec[i]) +
	static_cast<scalar_t>(gout_val_vec[i]);
	grad_vec[i] += static_cast<scalar_t>(gout_val_vec[i]);