bpo-46072: Add detailed failure stats for BINARY_OP

[brandtbucher]

On further reflection, I think it's definitely possible to get our current ~80% hit rate up to ~90%. These detailed failure stats will help guide my work on this.

Example full pyperformance run with this patch applied:

Kind	Count	Ratio
unquickened	7837567	0.1%
deferred	1187894724	19.8%
deopt	600612	0.0%
hit	4756557559	79.5%
miss	32075102	0.5%

Specialization attempts

	Count	Ratio
Success	948284	5.0%
Failure	18123381	95.0%

Failure kind	Count	Ratio
and int	2841285	15.7%
rshift	1964549	10.8%
lshift	1690147	9.3%
xor	1608323	8.9%
remainder	1574409	8.7%
add other	1411333	7.8%
true divide different types	1272004	7.0%
floor divide	993317	5.5%
true divide float	855450	4.7%
multiply different types	806998	4.5%
subtract different types	782979	4.3%
subtract other	722394	4.0%
power	598295	3.3%
or	345072	1.9%
add different types	326248	1.8%
multiply other	165969	0.9%
and other	140198	0.8%
true divide other	17550	0.1%
and different types	6861	0.0%

Note that "add other" and "multiply different types" also correspond more specifically to sequence concatenation and repetition, respectively. Similarly, "remainder" probably contains many examples of old-style string formatting.

I figure it's not worth the clutter of breaking out specific stats for these non-numeric cases yet, since their proportion of the total failures is still pretty low (and the benefit of specialization seems slim).

https://bugs.python.org/issue46072

[markshannon]

This looks like it adds some overhead when not gathering stats.
The way we usually avoid that is to put a function call into the macro, i.e:

default: SPECIALIZATION_FAIL(BINARY_OP, binary_op_failure_kind(lhs, rhs));