Small fix for "gradient not yet implemented" error

[j-towns]

There are various (admittedly quite obscure) situations where Numpy might throw a keyerror, and it could be incorrectly caught by this try ... except block, which would obscure the correct error. This slight rearrangement should prevent that from happening.

[j-towns]

A dumb example of what I'm talking about:

from autograd.core import primitive
from autograd import grad
import numpy as np

@primitive
def sin(x):
    return np.sin(x)

def sin_vjp(g, ans, vs, gvs, x):
    # generate a keyerror
    d = dict()
    d['a']

sin.defvjp(sin_vjp)

grad(sin)(1.)

seems to throw three exceptions, two of which are

NotImplementedError: Gradient of sin w.r.t. arg number 0 not yet implemented.