[XPU][Fix] Fix large maxpool index

[Stonepia]

This is to fix the pytorch/pytorch#167253 . It does the following:

1. Use index_t instead of int and dispatch kernels accordingly. (follows [CUDA] Large max pool fix pytorch/pytorch#167427)
2. Use NHWC when output > INT_MAX (follows cuda max_pool2d: switch to NHWC when output > INT_MAX to avoid overflow pytorch/pytorch#167322)
3. Change other related dtype (like num_wg to index_t to avoid overflow.

Details

Test case:

x = torch.zeros(74, 32, 30090, 81, device=torch.device("xpu"), dtype=torch.bfloat16)
torch.nn.functional.max_pool2d(x, kernel_size=(1,2), stride=(1,2), ceil_mode=False, padding=0)

It will throw the error:

[MaxPool2d] Input shape: [74, 32, 30090, 81] output: [74, 32, 30090, 40]
[MaxPool2d] Strides: n=77993280 c=1 h=2592 w=32
[MaxPool2d] Memory format: ChannelsLast
[MaxPool2d Forward] ChannelsLast path: numBatch=74 numPlane=32 inputH=30090 inputW=81 outputH=30090 outputW=40 index_t=int64
[MaxPool2d Forward] Using vec_size=1 num_wg=-72057583935701024
Segmentation fault from GPU at 0xff00000c04e33000, ctx_id: 1 (CCS) type: 0 (NotPresent), level: 1 (PDE), access: 0 (Read), banned: 1, aborting.
Segmentation fault from GPU at 0xff00000c04e33000, ctx_id: 1 (CCS) type: 0 (NotPresent), level: 1 (PDE), access: 0 (Read), banned: 1, aborting.
Abort was called at 279 line in file:
./shared/source/os_interface/linux/drm_neo.cpp
[1]    77805 IOT instruction (core dumped)  python

From the above code, the num_wg is overflow to negative, thus caused segfault.

[Copilot]

Pull Request Overview

This PR fixes integer overflow issues in XPU max pooling operations when handling large tensor sizes. The fix addresses a segmentation fault that occurred when output tensors exceeded INT_MAX by introducing index type templating and automatic memory format selection.

Key Changes:

• Introduced index_t template parameter (int32_t or int64_t) for kernel functors and functions to handle both small and large tensor sizes
• Added validation functions can_use_int32_nhwc and can_use_int32_nchw to determine when int32 is safe to use
• Automatically switches to ChannelsLast memory format when contiguous format would exceed int32 limits

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[EikanWang]

LGTM. Please address the copilot's comments.

[EikanWang]

Pls. fix the code style.

[Stonepia]

Thanks! Added in 474ac9d

[github-actions]

Performance outliers, please check!

• 🟡 [80%, 90%), may be fluctuations

Category	Model	Target vs. Baseline [Eager]	Target vs. Baseline [Inductor]
torchbench_bfloat16_training	resnext50_32x4d	0.934002	0.822647
torchbench_bfloat16_training	squeezenet1_1	1.021592	0.832874