Improved generics - Memloaders

[ChillFish8]

This is a small design doc around a change to the way we load memory and registers.

Problem

Currently, we have separate methods for performing vector x vector operations and vector x value operations, but this has a few issues:

• Requires a separate function for each combination (vector_x_value, value_x_value, vector_x_vector)
  • Currently value_x_vector ops are not implemented, but we do need them :(
• Harder to maintain across changes and inflates code-base.
• Requires that slice sizes are both the same, which artificially limits us to pre-broadcasting data before being able to correctly operate on it.

The initial trait idea

To skirt around these issues, the plan is to create a trait that represents a state machine, it knows the start and end bounds of its memory, the end result buffer size, and how it wants to load that data into our SIMD registers.

• This approach potentially allows us at a later stage to implement broadcasting without needlessly allocating, along with supporting strides, etc...
• Allows us to provide a single generic function that can take any pattern of op(value, value) , op(value, vector), op(vector, value), op(vector, vector)

[ChillFish8]

Last night we got #10 in which adds a base of the MemLoader system and fixes all the refactoring associated with that.

Now we should be in a good position to look at broadcasting and more advanced projections.