Classify specialization failures. Provides more useful stats, with lower overhead.

[markshannon]

This PR classifies specializes failures, so that we can get numbers on specialization failures.
This replaces the previous dictionary of unstructured data with an array of numbers, which can be collated using the
summarize_specialization_stats.py script.

The downside is that interpretation of the data now needs cross referencing with the list of SPEC_FAIL #defines.
IMO, being able to gather precise numerical values is easily worth that minor inconvenience.

[markshannon]

We should definitely turn on stats for debug builds, to be able to test it.
It would also make gathering stats a bit easier.

But that's for another PR.

[Fidget-Spinner]

I'll miss the old way a little -- it gave me an inkling of the exact attributes of types which were failing. However, this PR will definitely make collecting overall stats waaay better and the output seems less cluttered. Thanks Mark.

[Fidget-Spinner]

These numbers are repeated from above. Or is that intentional?

[markshannon]

Yes, they can overlap with each other. They shouldn't overlap with the common values, though.

[markshannon]

I'll miss the old way a little -- it gave me an inkling of the exact attributes of types which were failing.

When experimenting, it should be easy enough to change the set of SPEC_FAIL... values to add any additional ones that you want.
For example, I've just been using this branch to gather stats on LOAD_METHOD.