Unify GILGuard and Python

[konstin]

Currently, there a two types: GILGuard and Python, where GILGuard carries data and has a Drop implementation, while Python is a zero sized marked type. This leads to the following boilerplate all over the code:

let gil = Python::acquire_gil();
let py = gil.python();

We should unify those two types, so that we can just write

let gil = GIL::acquire();

and then use the gil just like the py now.

[fafhrd91]

I don’t think this is possible. Also acquire is rare for extensions

[joar]

Just to add a datapoint, I use acquire quite a lot in https://github.com/joar/rust-csv-py

[kngwyu]

@joar
Since acquiring gil can be overhead(though it's not so large), I recommend to use self.token.py() when it's possible(e.g. here).

[kngwyu]

If we remove Python<'p>(PhantomData<&'p GILGuard>), then we have to pass &'py GILGuard to functions, instead of Python<'p>.
So... what different does it make?
And, how about PyToken?
I have no idea to replace it with &GILGuard, without lifetime problem.

[konstin]

I'd rather like to remove PyToken completely (see #94).

For the GILGuard vs. Python story I thought about using things like Deref or a Python type that hides the GILGuard internally, but I haven't found a good design yet.

[kngwyu]

I see.
Yeah I feel like PyToken is not 'the best' approach, but have no better ides now 🤔

[davidhewitt]

I think that most users will use the new Python::with_gil API, so I'm not sure that any changes here are really necessary.

As a result I'm going to close this issue, but please shout if anyone thinks this still needs to happen and we can reopen.