Force specialization of dimension_mismatch_fail() on type argument.

This should improve GPU-compatibility.
Original discussion: https://discourse.julialang.org/t/problems-with-linearalgebra-functions-within-kernelabstractions-and-cuda/97566/5

src/convert.jl

@@ -192,7 +192,7 @@ AbstractArray{T,N}(sa::StaticArray{S,U,N}) where {S,T,U,N} = similar_type(typeof
 # Constructing a Tuple from a StaticArray
 @inline Tuple(a::StaticArray) = unroll_tuple(a, Length(a))
 
-@noinline function dimension_mismatch_fail(SA::Type, a::AbstractArray)
+@noinline function dimension_mismatch_fail(::Type{SA}, a::AbstractArray) where {SA <: StaticArray}
     throw(DimensionMismatch("expected input array of length $(length(SA)), got length $(length(a))"))
 end