#4790 zeroequiv should be more robust

[David Scherfgen]

When zeroequiv cannot symbolically show that an expression is non-zero, it uses numerical evaluation at a single random point between 0.1 and 0.9 and compares that to a threshold. The function could be improved a lot:

• Catch errors like division by zero, log(0), floating point overflow, ...
• Evaluate at multiple random points sampled from "all" real numbers, not just between 0.1 and 0.9
• Use relative, not absolute tolerance, based on the magnitude of sub-terms or derivatives