Add Insum Lowerer

[TheRealMichaelWang]

Added a Einsum to Indirect-Einsum (Insum) Lowerer. Additionally the following were also added:

• A GetAttribute EinsumExpr AST node. It gets an attribute/property of a tensor (i.e. the element or coordinate array of a COO tensor)
• A sparse tensor that inherits from eager tensor, with the corresponding FType. It is implemented in basic COO form.

The Insum Lowerer itself essentially does the following in order to convert an Einsum with a single sparse tensor to an Insum

• Build a tensor T whose elements store the number of non-zero elements in each corresponding fiber of the sparse tensor that is reduced
• Check if each reduced fiber in the sparse tensor is all non-zero or not by checking T. We then initialize the initial reduction values of our output tensor based on whether the corresponding reduced fiber is all non-zero or not.
  • If all elements in the reduced fiber of the sparse tensor are non-zero we can safely use the default initial value for the einsum's reduction operation
  • If there exist any zeros in the reduced fiber of the sparse tensor, we do the following.
    • We rewrite the arg/pointwise expression of the Einsum where we assume that all the accesses to the sparse tensor's elements are zero. We call this AST EinsumExpr, "ZeroArg".
    • We initialize the initial reduction value of our output tensor to the reduction of the default initial reduction value and the "ZeroArg" we calculated in the previous step.
    • In doing so we make an assumption about the reduction operator. For clarity let the reduction operator be called f, let the default initial value for f be called i, and let the array of elements being reduced called s. The core assumption we make is that if s has any zero elements, reduce(f, s) = reduce(f(i, 0), reduce(f, nonzero(s))). Also note that 0 is not necessarily the only value used in f(i, 0), but we still make a similar assumption.

[willow-ahrens]

I'm adding my review to remove the request, we've shifted our focus temporarily to the right-hand-side insum PR. Feel free to re-request once the other PR has been merged into main and integrated into this one and this one is ready.