Matmul implementation

[adeeconometrics]

Features of MatMulExpr operations:

• [x] Interopt with expression templates
• [x] Implement template predicates for matmul validity
• [x] Recursive expr tree matmul(expr, expr)
• [ ] Fuse arbitrary size matrix
• [ ] Add matrix algebra simplifications

[adeeconometrics]

• MatMulExpr API expects Expr to have Expr::rows() and Expr::cols()
• Matmul GEMM basic implementation can be found in MeshLines/matmul
• Compiler optimizations in Clang15 on M2 seem to work better for vanilla GEMM over blocked_matmul, threaded_blocked_gemm, async_gemm, and GEMM with NEON intrinsics. More to be explore on rewriting more efficient MatMul Expr

[adeeconometrics]

Variadic matmul is a contentious path forward as it is likely to degrade performance.

Take this step when one of the following conditions are satisfied:

• Until a performance gain is obtained in lazy matmul implementation
• Until a promising algorithm for fusing operations is found