Round the epilogue offset to positive values in cunn_SoftMaxForward

[xinyazhang]

This fixes OOB memory access for followng code

import torch
qk = torch.randn((9,1017), dtype=torch.float64, device='cuda')
smqk = torch.softmax(qk, dim=-1)

Correctness can be confirmed with:

import torch
import numpy as np
from scipy.special import softmax

qk = torch.randn((9,1017), dtype=torch.float64, device='cuda')
nqk = qk.cpu().numpy()
smqk = torch.softmax(qk, dim=-1)
nsmqk = smqk.cpu().numpy()
smnqk = softmax(nqk, axis=-1)

print(f'{np.allclose(smnqk, nsmqk)}')

This is ported from upstream PR pytorch#154634

[pruthvistony]

@xinyazhang ,
Can this PR be closed?

[xinyazhang]

Duplicated with #2247