Specialize LinearAlgebra.BLAS.dot for strided vectors of floats.

[ghost]

Fixes #37767.

(tested using openblas64)

[ViralBShah]

Welcome @dagbr! Just holding this one while we try to get #39455 merged. I think this shouldn't lead to a conflict, but there's a lot of linear algebra stuff changing simultaneously, and just being cautious for now.

[dkarrasch]

Out of curiosity, what exactly was the issue and how did you resolve it?

[ghost]

I have amended my commit to remove a misleading comment.
@dkarrasch
It seems StridedVector{Float64}'s are treated without taking into account views with negative step size.
A StridedVector can be a SubArray arising from a view like v = view([1.0,2.0,3.0], 3:-1:2) with negative stride.
When calling OpenBLAS, we should pass a pointer to first element in memory, which is at pointer(v,length(v)) if the stride is negative.
My guess is that, in the case of a SubArray, the conversion to Ptr{Float64} yielded some other pointer, I think it could be Base.unsafe_convert(Ptr{Float64},v) which is equal to pointer(v,1) (at least when I played with it in the REPL).
I think this can happen because the wrappers in blas.jl for the dot functions from BLAS take any AbstractArray.

[ViralBShah]

#39455 is merged. So that is no longer blocking this PR.