Fix some B023 false alarms

[jakkdl]

Picking up from main...Zac-HD:flake8-bugbear:B023-false-alarms
Partially fixes #269

Fixes some common false alarms where a function is immediately consumed, as an argument to filter or reduce, as a keyword argument to key= and as a return value.

TODO: I'm not sure how to handle more complex return cases, see the last test cases in b023.py.

I'm also not sure if this solution is too generous, it currently marks all direct child nodes to calls to filter&reduce as safe - but @Zac-HD's comment implies it should maybe be stricter:

Unfortunately it's more difficult to avoid this false alarm than you might think: for example "lambdas are OK if they're the first argument in a call to filter()" would miss bugs where the filter iterable (and therefore lambda) isn't immediately evaluated.

I'm not super read up on the problem and didn't take time to fully understand all comments with false alarms in the original issue, so would love some eyes on this. But it fixes all test cases @Zac-HD added at least.

[Zac-HD]

Looks good to me - thanks @jakkdl!

cc @cooperlees to merge?

[cooperlees]

Thanks. Lots going on here I don't get but I'll trust you two! Tests look good.

[Zac-HD]

I'm also not sure if this solution is too generous, it currently marks all direct child nodes to calls to filter&reduce as safe - but @Zac-HD's comment implies it should maybe be stricter:

Unfortunately it's more difficult to avoid this false alarm than you might think: for example "lambdas are OK if they're the first argument in a call to filter()" would miss bugs where the filter iterable (and therefore lambda) isn't immediately evaluated.

In short, there was enough chatter in that thread to make it clear the false-alarm rate was unacceptably high, and if we miss some real alarms that's basically just the natural downside of minimally-resourced open source projects! Better to catch most bugs without annoying people than have a tool that nobody uses...