Level 1 BLAS functions (except for dot) dont handle inputs with negetive stride correctly

[N5N3]

Similar to JuliaLang/julia#30765.
An eazy example:

julia> a = collect(1.0:5.0)
5-element Vector{Float64}:
 1.0
 2.0
 3.0
 4.0
 5.0

julia> v = view(a, 3:-1:1)
3-element view(::Vector{Float64}, 3:-1:1) with eltype Float64:
 3.0
 2.0
 1.0

julia> BLAS.nrm2(v)
7.0710678118654755

julia> BLAS.nrm2(collect(v))
3.7416573867739413

[oscardssmith]

is it worth keeping level one blas around? I don't think it's faster than regular broadcasting...

[N5N3]

At present, norm(::StridedVector, 1) and norm(::StridedVector, 2) call BLAS.asum and BLAS.nrm2 for long enough inputs (and return wrong result for negetive stride).
So we'd better to ensure correctness?

On the other hand, we have JuliaLang/julia#41513 for gemv, and JuliaLang/julia#39751 for dot.
There seems no reason we dont add negative stride support to other functions, (if they return correct results)

[sostock]

Duplicate of #813.