Run benchmark analyses even after failure

[Chris Warburton]

In particular, we need to know the number of wanted conjectures for
*all* data points, including those which failed. This is easy enough,
since it only depends on the sample, which we know whether or not the
exploration succeeded. We can supply '[]' to the analyser's stdin in the
case of a failure, which will give us 0 recall, undefined precision and
a list of all the wanted conjectures.

The reason we need to know this is to work out the standard deviation of
our precisions and recalls.