AbstractArray -> StridedArray

[willtebbutt]

#21 essentially points out that we have the same problem that we see regularly in the AD world, where a sensible implementation of flatten for e.g. a sparse array or Fill is not the same as a sensible implementation for e.g. a StridedArray.

Should be a quick fix. We probably need to copy over + tweak the code for to_vec for general structs so that flatten gives you the struct interpretation of an array by default.

Disclaimer: I'm assuming thay if a user writes Fill(5.0, 10) then they want to constrain this vector to be a Fill throughout e.g. optimisation, and that if this weren't the case said user would write fill(5.0, 10).