Bring our TokenTree model closer to rustc's

[Veykril]

Our model is kind of arbitrary and less strict, the syntax node <-> TokenTree conversion is also messy and brittle, likewise our MBE infra will benefit from a more concise model. Lastly, rustc uses this one TokenStream model for the parser, be it from macro inputs/outputs or general lexing where as we have a very confusing and inefficient setup to make this work for our parser. Ideally we should do a similar move and switch to a unified input there which ought to aid perf quite a bit when transcribing MBE macros.

[bors]

☔ The latest upstream changes (presumably #17994) made this pull request unmergeable. Please resolve the merge conflicts.

[ChayimFriedman2]

Switching to a unified representation is problematic because we need syntax trees to be lossless but tt must not be lossless - to help incrementality, but I do have a sketch of idea.

[Veykril]

Fyi I am currently bringing this PR back on track (module the mbe crate, will likely rewrite it on top of this change), as changing the token tree representation here is still beneficial either way. The unified input thing is a different story, unrelated to this PR in my eyes.

[Veykril]

I started work on a fresh branch for this but I now noticed that swapping this is quite the pain due to our parser setup. The parser needs a dire rewrite as well, the way it handles tokens it horrible for wrangling macros so this goes back to the input structure of the parser needing to be changed