Support Einsum operator

[robertknight]

This is used by Segment Anything amongst other models.

• Add initial implementation for cases that can be handled as a matmul (+ transpose) (Add partial implementation of Einsum and Segment Anything example #295)
• Add initial support for other equations, except for ellipsis and repeated dimensions (Expand supported equations in Einsum #297)
• Support ellipsis to control broadcasting (Support ellipsis for broadcasting in Einsum equations #302)
• Handle reducing over multiple dimensions efficiently (ie. by reducing to matmul or similar) (Use matmul to reduce multiple dimensions in Einsum more efficiently #299)
• Handle more than two input terms efficiently if reductions are involved (Support using matmul for Einsum equations with 3 or more inputs #300)
• Handle repeated dimensions within a single input. term. Per the numpy.einsum docs, this takes the diagonal (Take diagonals when labels are repeated in an Einsum input term #303)
• Error if repeated dimensions appear in an output term (Error if an Einsum term contains repeated labels or ellipsis #301)