Sparsity support

[chhzh123]

This thread aims to add sparsity support to Allo, starting from the 2:4 pattern.

Reference: https://developer.nvidia.com/blog/accelerating-inference-with-sparsity-using-ampere-and-tensorrt/#sparse_tensor_cores_accelerate_24_fine-grained_structured_sparsity

[CynyuS]

Ran the github systolic array examples.
Read the theoretical background for systolic arrays.
Started implementing a naive solution to multiply one sparse and one dense matrix.
Currently researching optimizations/parallelizations to simplify calculations for the matrix multiplication of a sparse and dense matrix, utiliziting the fact that the sparse matrix is half zeros.

[CynyuS]

In the process of implementing the kernel ran into some setup errors in Allo. Is currently in the debugging process to re-setup allo.

[CynyuS]

Created basic systolic array algorithm. Added if statement to check if sparse element is zero, and if it is, just pass c on.
Working on advanced algorithm with an extra fifo input of the indices.
Next steps: Optimize simple algorithm and perform speed tests.