Split mp_binary_op into inlinable and non-inlinable variants.

[ltratt]

mp's normal mp_binary_op inlines iterator comparisons for contains: the resultant while loop prevents mp_binary_op from being inlined in a trace. This commit breaks out the arbitrary-length while loop into a separate function: normal binary operations are now inlined, but list (etc.) comparisons aren't. The one exception is tuples: those are inlined (and therefore unrolled), because they are almost always small and fixed in size.

This speeds nearly everything up, if only by a bit:

BigLoop/ykmp/1000000000  227746 ±  976  136731 ±    5   0.60  39.96% faster
Sieve/ykmp/3000           19497 ±   59   15379 ±   79   0.79  21.12% faster
Queens/ykmp/1000          70252 ±  557   66450 ±  111   0.95  5.41% faster
List/ykmp/1500            62093 ±  626   58845 ± 1500   0.95  5.23% faster
Towers/ykmp/600          153793 ±  312  145922 ±  541   0.95  5.12% faster
Richards/ykmp/100        355874 ±  361  337834 ±  693   0.95  5.07% faster
Bounce/ykmp/1500         122221 ±  811  117012 ± 1106   0.96  4.26% faster
Permute/ykmp/1000        133177 ±  171  128087 ±  329   0.96  3.82% faster
CD/ykmp/250              209803 ± 1146  202455 ±  381   0.96  3.50% faster
NBody/ykmp/250000        112451 ±   92  113991 ± 1165   1.01  1.37% slower
Storage/ykmp/1000        112431 ±31069   88203 ±  293   0.78  indistinguishable
Mandelbrot/ykmp/500       44276 ± 1156   42130 ± 1316   0.95  indistinguishable

[Pavel-Durov]

Looks like test failed:

11:49:04 2 tests failed: basics/int_big_mul.py thread/thread_lock4_intbig.py

[ltratt]

Yes, currently investigating. This is probably a deopt bug but not 100% sure yet.