What gradient should the `**` operator produce for the exponent when the base is negative?

[axch]

The "correct" answer is that exponentiation with negative base and real-valued exponent is undefined, so crash. However, a relatively common use case of ** is to spell polynomials, as with x**3. In this case, the gradient of ** should clearly be [3x**2, 0], even for negative x, as the exponent is intended to be a non-varying integer.

If we had first-class integers (#436, #570) in Venture I would be comfortable defining the gradient of ** to be generic in the argument type, and produce the latter answer if the exponent came in as an integer. That policy would also suggest doing the same thing for other type-generic numerical operations, and/or teaching integers to accept "gradient" steps but not move.

[riastradh-probcomp]

In commit 09876c9f75f512b8c2887d0ba5258414f09b81d9 I replaced math.log by np.log in the code to compute the gradient of pow, so that for the partial wrt the exponent it gives NaN instead of raising an exception when the base is a negative real. I have not confirmed that this change doesn't have bad consequences, but it does seem to get around the immediate problem I had with gradient methods, which now give a sensible result.

[axch]

I think actually solves this.