Activation support for GroupedTensor: BF16→BF16 version

[ptrendx]

Develop support for applying activations to GroupedTensor data in BF16 format, with the optional amax output.

[vthumbe1503]

Fused Kernels for Scaled Activations needed by MOE were added as part of #3132

And the tex pytorch APIs are added as part #2988

Currently they consume BF16 Tensor instead of GroupedTensor. The above scaled Activation APIs need to be extended to be used with GroupedTensor so that we dont do wasted computations on the padded portions of GroupedTensor