[FEATURE] Int8 kernels for Sparsity

[vkkhare]

Describe the feature request
Since quantization techniques are orthogonal to sparsity, we should be able to leverage the benefits of both and stack them together.

Describe the solution you'd like

We have similar dtype templates in cuda which we need to replicate for CPU and instruction sets like AVX

template <>
__global__ void sparse_mlp_combined_cuda_kernel<float>(...)