[Prim] Implement group_norm_backward (#84037)
Test plan: CI, i.e. `python3 test_decomp.py -v -k test_comprehensive_nn_functional_group_norm` plus:
```
#!/usr/bin/env python3.8
import torch
func = torch.ops.aten.native_group_norm_backward.default
decomp = torch._decomp.decomposition_table[func]
for args in (
(torch.rand(1, 6, 3), torch.rand(1, 6, 3), torch.rand(1, 2), torch.rand(1, 2), torch.rand(6), 1, 6, 3, 2, [True, True, True]),
(torch.rand(64, 768, 7, 7), torch.rand(64, 768, 7, 7), torch.rand(64, 1), torch.rand(64, 1), torch.rand(768), 64, 768, 49, 1, [True, True, True])):
nrc=func(*args)
drc=decomp(*args)
for i in range(len(nrc)):
print(i, torch.max(nrc[i]-drc[i]))
print(all(torch.allclose(x, y) for (x, y) in zip(nrc, drc)))
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/84037
Approved by: https://github.com/Chillee, https://github.com/ngimel
