Drop subtyping of DenseArray

[andyferris]

I had originally defined

abstract StaticArray{T,N} <: DenseArray{T,N}

with the thought that a lot of functionality could be shared with Base.LinAlg functions, such as automatically calling out to BLAS functions and so-on.

In the end, I don't feel this has been as successful as I hoped. Most of the functions lack type-stability because Base makes a habit of calling similar (with the same size). So not only do users not get, e.g., an SMatrix back (they get an MMatrix), the code inside the functions (and sometimes the output) is frequently not type-stable and is probably much slower than just using Base.Array.

Further, most of the decompositions actually fail. Only eig seems to work well, and it is also not type stable internally.

And finally, we already have StaticArrays which are not dense, like Rotation and probably in the future things like static unit ranges and so-on.

So the proposal is to change to

abstract StaticArray{T,N} <: AbstractArray{T,N}

and to therefore drop some free functionality (and implement things ourselves).

[c42f]

👍 from me, these reasons seem compelling.

[andyferris]

The question is then whether to do this separately from an attempt to add the size as a type parameter.

[andyferris]

Though I'm leaning towards traits for that.

[c42f]

The options for putting the size as a type parameter seem somewhat poor without better compiler support. I think it'd be worth trying a traits-based version.

[andyferris]

See #57