[FEAT] Allow training m without a blocking rule with a sample of the input records

[RobinL]

In some cases it might be easier to train m by taking a sample of input records and training on the cartesian product. This should work fine so long as the sample contains 'enough' matches such that there are a decent number of examples in the pairwise predictions of each comparison level