bpo-45885: Don't un-adapt COMPARE_OP when collecting stats

[brandtbucher]

The current stats obscure the fact that "not followed by conditional jump" accounts for a significant portion of possible COMPARE_OP hits. This patch takes the same approach that BINARY_OP uses of only un-adapting the instruction when stats aren't being collected.

New stats:

COMPARE_OP

specialization stats for COMPARE_OP family

Kind	Count	Ratio
unquickened	7362156	0.4%
specialization.deferred	288102012	17.2%
specialization.deopt	15642	0.0%
hit	1374191464	82.3%
miss	959440	0.1%

Specialization attempts

	Count	Ratio
Success	373171	7.4%
Failure	4650804	92.6%

Failure kind	Count	Ratio
not followed by cond jump	2647332	56.9%
float long	665643	14.3%
set	580212	12.5%
different types	151788	3.3%
bool	137583	3.0%
other	127248	2.7%
tuple	120230	2.6%
big int	101401	2.2%
bytes	52839	1.1%
baseobject	42585	0.9%
list	23036	0.5%
string	706	0.0%
long float	201	0.0%

https://bugs.python.org/issue45885

[sweeneyde]

Nice -- this puts this failure kind on equal footing with the other failure kinds so the stats are more honest.

[markshannon]

👍