Type-stability fixes for matmul

[timholy]

This one is quite funny: we have

julia> v = [1,2]
2-element Array{Int64,1}:                                                                                                                                                          
 1                                                                                                                                                                                 
 2

julia> using Base.Test

julia> @inferred(v*v')
2x2 Array{Int64,2}:
 1  2
 2  4

but also

julia> C = Array(Int, 2, 2)
2x2 Array{Int64,2}:
 140643182283632  140634494951504
 140634541475488                0

julia> @inferred(A_mul_Bc!(C, v, v))
ERROR: return type Array{Int64,2} does not match inferred return type Any
 in error at ./error.jl:21

The reason turned out to be a bit subtle:

julia> Base.LinAlg.matmul2x2!(C, 'N', 'C', v, v)
ERROR: MethodError: `matmul2x2!` has no method matching matmul2x2!(::Array{Int64,2}, ::Char, ::Char, ::Array{Int64,1}, ::Array{Int64,1})
Closest candidates are:
  matmul2x2!{T,S,R}(::AbstractArray{R,2}, ::Any, ::Any, ::AbstractArray{T,2}, ::AbstractArray{S,2})

It seems that when no method exists, inference is returning Any rather than Union{} for the return type?

So the underlying cause is the branch for 2x2 or 3x3 matrices, the fact that the corresponding functions are defined only for AbstractMatrix inputs, and the fact that generic_matmatmul! can take inputs which are not AbstractMatrix objects.

Also fixes another type instability in copy! (it was returning Union{Bool,typeof(B)} which probably has no performance impact, but we might as well fix it.

[timholy]

Might want to wait a bit on merging this to see whether there's an easy, backportable fix for #14336.

[timholy]

Looks like #14336 might have to wait on #265, and we all know what that means. I'll merge this, and it also should be backported.