finite_difference routines temporarily mutate inputs

tl;dr The finite_difference methods for the gradient and jacobian (but not hessian) temporarily mutate the input `x` vector in order to compute the derivatives. I think this is unwise since there is no way to know inside these functions if the input `f` function holds a reference to `x` and the user will not suspect this will happen. A copy of `x` should be made to make these routines safer.

More details for those interested:

This recently happened to me and it was difficult to track down. It occurred because I called the curve_fit routine from LsqFit.jl:

```
function curve_fit(model::Function, xpts, ydata, p0; kwargs...)
```

and inputted the same array for `ydata` and `p0`. I know this seems unusual but it makes sense for the problem I was solving (I can provide details if you want). This caused the fit to fail because `curve_fit` makes a function that includes `ydata` as a closure that is used in `finite_difference_jacobian!`. I don't see how a user of curve_fit would suspect that somewhere under the hood an input array will be mutated (even temporarily).

I see that the Calculus.jl routines are going to be superseded by the FiniteDiff.jl and this concern affects routines there as well (for computing the gradient and hessian; I don't see a jacobian routine there).

Thanks for your consideration.