macros: optimize generated code for #[derive(FromPyObject)]

[davidhewitt]

Inspired by #1966, I decided to take a look at the generated code for #[derive(FromPyObject)] for enums.

There seemed to be some pretty easy wins as well as fixes to error messages, so I just went ahead and made them. Mostly what I did is delay all string formatting until creation of the final error message, to avoid intermediate costs when the conversion still ultimately succeeds. To keep generated code smaller I also extracted this into a #[cold] helper function that gets re-used for all enum derive implementations.

Added a benchmark for a simple enum case; it shows quite significant wins:

enum_from_pyobject      time:   [536.93 ns 541.67 ns 547.81 ns]
                        change: [-69.528% -69.164% -68.793%] (p = 0.00 < 0.05)
                        Performance has improved.

... there's probably more that can be done here to improve the other generated code from this macro too; this is enough for today.