implement a softmax version of focal loss

[wyli]

Is your feature request related to a problem? Please describe.

The softmax version of focal loss is:

  FL(p_t) = -alpha * (1 - p_t)**gamma * log(p_t),

where p_i = exp(s_i) / sum_j exp(s_j), t is the target (ground truth) class, and
s_j is the unnormalized score for class j.

cf https://github.com/pytorch/pytorch/blob/main/modules/detectron/softmax_focal_loss_op.cc#L38-L43

the current implementation is hard-coded to use sigmoid

MONAI/monai/losses/focal_loss.py

Line 26 in e5cf74a

class FocalLoss(_Loss):

[qingpeng9802]

I can work on this issue and PR later if no one is working on this.