fix vec and diag for RowVector

[bkamins]

I propose to fix two issues with RowVector.

1. vec on RowVector can be made more efficient by defining a specialized method
2. currently diag on RowVector returns RowVector and it should return a Vector; in this PR this is fixed

The reason for fixing diag is the following situation:

julia> [2 2; 2 2] / [2]
ERROR: StackOverflowError:

whereas it should be [1.0 1.0; 1.0 1.0] as it is the result of reshape([2], 1, 1) \ [2 2; 2 2].

Additionally (I have not touched this because it is probably intentional - if not I propose to fix it):

@propagate_inbounds getindex(rowvec::RowVector, i) = transpose(rowvec.vec[i])

which means that indexing into RowVector by a vector produces RowVector. This is inconsistent with the general behavior that indexing into e.g. Matrix by a vector produces Vector, e.g.

reshape([1,2], 1, 2)[[1]]

is a Vector, but

[1,2]'[[1]]

is a RowVector.

[fredrikekre]

All of this seem to work on master?

julia> [2 2; 2 2] / [2]
2×2 Array{Float64,2}:
 1.0  1.0
 1.0  1.0

julia> reshape([1,2], 1, 2)[[1]]
1-element Array{Int64,1}:
 1

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

Also, RowVector will be gone after #25217

[bkamins]

@fredrikekre sorry for the mess - I keep forgetting that recent pace of improvements in Julia is so fast and checked the code on older version. Closing.

[Sacha0]

No worries @bkamins! Rather, much thanks for your wonderful recent efforts! :)