[ML] Memory usage of hierarchical results normalizer is not accounted for

[droberts195]

We do not account for the memory size of the hierarchical results normalizer.

When it was first introduced it was fairly close to being a fixed size object, so could reasonably be considered part of the "process overhead" - fixed data structures for which we add 10MB to the required memory of each job.

However, now that we do per partition scoring it is no longer reasonable to consider the normalizer part of the fixed overhead. For a job with a large number of partition field values it can be a non-negligible size, and this memory should be properly accounted for.