Update LayerNorm with Welford online algorithm

[min-jean-cho]

Welford online algorithm improves the numerical stability of variance computation compared to naive or two-pass algorithm.

Simple test case:

import torch
import torch.nn as nn

B, D = 1, 5
x = torch.tensor([[1e15, 1e15 + 1, 1e15 + 2, 1e15 + 3, 1e15 + 4]], dtype=torch.float32).to("xpu")
layernorm = nn.LayerNorm(D, elementwise_affine=False)
y = layernorm(x)
print("LayerNorm Output:\n", y)

Output now (welford):

tensor([[0., 0., 0., 0., 0.]], device='xpu:0')

Output before (two-pass):

tensor([[nan, nan, nan, nan, nan]], device='xpu:0')

[xytintel]

Please add test case

[dvrogozh]

Build fails against latest pytorch with:

/home/dvrogozh/git/pytorch/pytorch/third_party/torch-xpu-ops/src/ATen/native/xpu/sycl/LayerNormKernels.cpp:365:37: error: use of undeclared identifier 'get_group_reduce_group_size'
  365 |         sycl_local_acc_t<float>(2 * get_group_reduce_group_size(sg_size_), cgh);
      |                                     ^
/home/dvrogozh/git/pytorch/pytorch/third_party/torch-xpu-ops/src/ATen/native/xpu/sycl/LayerNormKernels.cpp:367:33: error: use of undeclared identifier 'get_group_reduce_group_size'
  367 |         sycl_local_acc_t<float>(get_group_reduce_group_size(sg_size_), cgh);
      |                                 ^
2 errors generated.

Please, fix. I would like to verify PR against pytorch/pytorch#141642.

[xytintel]

@min-jean-cho Please resolve case failure

[dvrogozh]

Works for me to resolve pytorch/pytorch#141642. Please, update PR title (drop "[WIP][Test]") if ready.