Fix memory leaks found by memleak_hawaii3.py

[mdboom]

Sorry I couldn't sort this one out before 1.5.0.

[WeatherGod]

The changes here make sense. Are there other places where convert_to_string was (mis)-used?

[mdboom]

No -- convert_to_string is only called from one place: Py_convert_to_string.

[pelson]

Do you want to target the v1.5.x branch?

[mdboom]

Might as well target 1.5.x

EDIT: But we can just merge to master and backport after.

[WeatherGod]

backported to v1.5.x as a450f99

[efiring]

@mdboom Thanks very much for tracking this down. Out of curiosity: do you have a simple explanation for why the leak was so much larger on Python 3 than 2?

[mdboom]

Unfortunately, no. But I was seeing a great deal of variation in the result from the memleak_hawaii3.py script -- I think it was actually calculating the average increase incorrectly, so it may well be that the numbers were just bogus. (It was calculating the difference between the memory usage in the middle and the end and dividing by the steps. That only makes sense if the range is monotonically increasing. If it's jagged -- which it is -- you can get wildly incorrect numbers. See #5360).