Repeted evaluation of `fx0` when computing forward differences for gradients

[junyuan-chen]

For f that maps a vector of length N to a scalar, I am expecting f to be evaluated for only N+1 times with fdtype == Val(:forward). What actually happens is that the same fx0 = f(x) is evaluated for N times as iterating across all elements of the input vector. Hence, f is called 2N times!

See the line below that should be pulled out of the loop:

FiniteDiff.jl/src/gradients.jl

Line 383 in 83ff37e

fx0 = f(x)