Fix type stability bug in gradient

cc @timholy

I don't know if this is the best possible solution to the problem, but
it does allow us to do things like gradient(itp, dual(2,1)) and get an
Array{DualNumbers{Float64},1} back.

Before, that incantation would throw an InexactError because it was
trying to fit a dual number with nonzero epsilon into an Array{Float64}.

Author: tomasaschan

src/b-splines/indexing.jl

@@ -72,10 +72,8 @@ end
 
 @generated function gradient{T,N,TCoefs,IT,GT,pad}(itp::BSplineInterpolation{T,N,TCoefs,IT,GT,pad}, xs...)
     n = count_interp_dims(IT, N)
-    quote
-        @which gradient!(Array(T, $n), itp, xs...)
-        gradient!(Array(T,$n), itp, xs...)
-    end
+    Tg = promote_type(T, xs...)
+    :(gradient!(Array($Tg,$n), itp, xs...))
 end
 
 @generated function gradient!{T,N}(g::AbstractVector, itp::BSplineInterpolation{T,N}, xs::Number...)