remove ambiguity in unsafe_convert with RefValue

[sjkelly]

Before:

julia> detect_ambiguities(Base, LinearAlgebra, StaticArrays)
2-element Array{Tuple{Method,Method},1}:
 (unsafe_convert(::Type{Ptr{T}}, a::Base.RefValue{SArray{S,T,D,L}}) where {S, T, D, L} in StaticArrays at /home/steve/.julia/dev/StaticArrays/src/SArray.jl:69, unsafe_convert(::Type{Ptr{Nothing}}, b::Base.RefValue{T}) where T in Base at refvalue.jl:30)
 (convert(::Type{SA}, a::AbstractArray) where SA<:(SArray{Tuple{},T,0,1} where T) in StaticArrays at /home/steve/.julia/dev/StaticArrays/src/Scalar.jl:13, convert(::Type{SA}, sa::StaticArray) where SA<:StaticArray in StaticArrays at /home/steve/.julia/dev/StaticArrays/src/convert.jl:10)

After:

julia> detect_ambiguities(Base, LinearAlgebra, StaticArrays)
1-element Array{Tuple{Method,Method},1}:
 (convert(::Type{SA}, a::AbstractArray) where SA<:(SArray{Tuple{},T,0,1} where T) in StaticArrays at /home/steve/.julia/dev/StaticArrays/src/Scalar.jl:13, convert(::Type{SA}, sa::StaticArray) where SA<:StaticArray in StaticArrays at /home/steve/.julia/dev/StaticArrays/src/convert.jl:10

The remaining ambiguity is covered by #774

[sjkelly]

The failure on nightly is reproducible locally. It also affects master.

[c42f]

It's mysterious to me that unsafe_convert(Ptr{Nothing}, Ref(SA[1,2])) works without this fix rather than throwing a MethodError. Maybe detect_ambiguities is not precise?

Relatedly, I do wonder whether it would be better to deprecate the associated cconvert, it's more explicit to have people manually wrap in a Ref if they need to pass an SVector to a ccall. But that's a separate issue.

In any case, thanks for fixing this. Let's merge it!

[c42f]

Ah I got a little trigger happy; we should also have reduced the ambiguity counter in the tests to go with this improvement - could I bother you to do that as well?

[c42f]

Relatedly, I do wonder whether it would be better to deprecate the associated cconvert

I started opening an issue about this, but in the process realized that the current behavior is pretty reasonable. So move on, nothing to see here :)

[sjkelly]

It still doesn't really make sense to me, but I am not well versed in the inference/matching mechanics. If I remember correctly the behaviour is the same on 1.0 and 1.1 as well.

Ah I got a little trigger happy; we should also have reduced the ambiguity counter in the tests to go with this improvement - could I bother you to do that as well?

I'll test on prior Julias and get bounds in a different PR. I think 1.0 still has 2 or 3 remaining. So maybe have tighter bounds for 1.4+.

[c42f]

I think 1.0 still has 2 or 3 remaining. So maybe have tighter bounds for 1.4+.

Perfect, much appreciated!