Regression in import time

[Dreamsorcerer]

In aiohttp we have a regression test for the import time of the library.
When using idna <=3.6, the test is passing, but with idna 3.7+ the test fails consistently, indicating a significant regression in import time.

Looking at the changes, the only obvious change is the massive amount of data added to idnadata.py (in joining_types). Maybe this could be lazy loaded, e.g. by putting the construction in a function with a @cache decorator, or by coming up with a more efficient data structure?

[kjd]

Will look into improving the performance here, patches are welcome!

I haven't looked into it, but some of the odd data structures we use are informed by Jython and a 64kb limit it has on such structures, which I suspect he recent changes to joining_types also transgresses. Need to find a way to test if that is also affected as part of looking into this.

[Dreamsorcerer]

I have no idea what joining_types actually does, so don't have any obvious suggestions on how to improve the data structure. If you want to go with the lazy load function, I can take a look at that later. Should be pretty trivial.