GH-117581: Specialize binary operators by refcount as well as type.

[markshannon]

This PR changes the specialization of binary operators for the current, fairly ad-hoc approach to a bit more principled approach of specializing by refcount and using table lookup to specialize by type.

Tier 1 performance is in the noise, maybe showing a slight slowdown.

I expect the advantages of this approach to show up when applied to BINARY_SUBSCR, COMPARE_OP and with the JIT.

It also opens the possibility of specializing binary operators for numpy arrays and the Decimal class.

• Issue: Specialize BINARY_OP by refcount and by type of operands #117581

[Fidget-Spinner]

Can you please document in the bytecodes.c file what these suffixes mean? Some of them are easy to see, but I don't know what's NN vs ND immediately.

[markshannon]

Done.

[markshannon]

Performance on both tier 1 and tier 2 is about 1% slower.
With deferred reference counting, there will be no need for specializing on refcount, so I'm deferring this until deferred reference counting is done.

[markshannon]

Earlier results had an issue with the telco benchmark, probably due to the decimal module falling back to the pure Python version.

Benchmarking on tier 1 shows the slowdown to be in the noise.
The significant results are a speedup of 7% for spectral_norm and a slowdown of 12% for nbody.
This is presumably due to more complete specialization for spectral_norm and additional overhead for the already well specialized nbody code.
We expect the advantage of better specialization to be important in tier 2, but the additional tier 1 overhead to be largely irrelevant there.

[markshannon]

Abandoning this as #128722 is handling the same problem and we have a better ideas for reducing refcounting overhead faster-cpython/ideas#700