MLP*: Doesn't respect input IDs

[weefuzzy]

The MLP client training code seems to assume that the input and output data / labels are identically ordered, and never checks that this is the case. Much of the time it is, but one can simply 'reorder' the entries of a Dataset – which should remain semantically equivalent because it's an unordered collection – and completely screw up one's training.

There's no particularly nice solution to this: we just have to traverse the ID list of one of the data sets and rearrange (a copy of) the other to match. Current fix in progress introduces a new flag (strict) to enable this, because it's not cheap.