Add LOAD_FAST__LOAD_ATTR_INSTANCE_VALUE combined instruction

[sweeneyde]

This pair accounts for 3.1% of all opcode pairs (the next runner-up is only 1.8%). Hopefully that would translate into about 3.1% of total instructions executed being LOAD_FAST__LOAD_ATTR_INSTANCE_VALUE, or a bit more if we make sure to prioritize it over LOAD_FAST__LOAD_FAST, LOAD_ATTR_INSTANCE_VALUE, which is now at 0.3% of all pairs.

For reference, current super/combined-instruction stats are:

LOAD_FAST__LOAD_FAST	3043493930	4.5%
STORE_FAST__LOAD_FAST	2993756763	4.5%
STORE_FAST__STORE_FAST	1917861514	2.9%
LOAD_FAST__LOAD_CONST	1553531288	2.3%
LOAD_CONST__LOAD_FAST	567105342	0.8%

LOAD_FAST__LOAD_ATTR_INSTANCE_VALUE can also probably elide an INCREF/DECREF pair for owner in the expression y = owner.attr, and avoid touching the stack until the final PUSH(result).

The super instruction can be added at specialization time, something like

    ... after specialize_dict_access ...
    if (_Py_OPCODE(*instr) == LOAD_ATTR_INSTANCE_VALUE) {
        if (_Py_OPCODE(instr[-1]) == LOAD_FAST) {
            instr[-1] = /* LOAD_FAST__LOAD_ATTR_INSTANCE_VALUE */;
            if (_Py_OPCODE(instr[-1]) == LOAD_FAST__LOAD_FAST) {
                instr[-2] = /* LOAD_FAST */;
            }
        }
    }

Upon deoptimization, it can deoptimize back to LOAD_FAST, without having had any effect.

[sweeneyde]

python/cpython#31484