Empty versions of `Span::start` and `Span::end` returned on stable compiler

[udoprog]

Introduced in #166.

With a stable compiler with the span-locations feature enabled, we now have a situation where Span::start and Span::end reports mocked values (LineColumn { line: 0, column: 0 }) (see code). This is understandable since the underlying function in the compiler is nightly-only behind a feature flag (proc_macro_span and rust-lang/rust#54725), but it was somewhat surprising to me that proc_macro2 falls back to reporting fake spans.

The surprising behavior can be mitigated by documenting in Span::start and Span::end that their values will be mocked on stable compilers until the relevant features are stabilized. But also, downstream projects which for whatever reason rely on accurate spans don't have a great avenue for erroring at compile time in case the compiler version is too old to support them*.

*: Span::unwrap could in principle be used, but would mean that the code no longer can run outside of a proc_macro.

This is also a potential future compatibility hazard if start and end were to be broken out into a separate feature flag (as discussed here) to ease the road towards future stabilization. So I'm tying it together here since I'm looking into if and how it can be done.

[chipsenkbeil]

I got a bit excited that this was available in stable because it would make hacks like proc-macro-faithful-display be available to stable as well. I didn't see anything mentioned in the documentation for Span::start or Span:end in proc_macro2 that the values would be mocked.

It makes total sense as I had no idea how proc_macro2 was going to do this in stable if it was to have parity with proc_macro. But a disclaimer in the method docs about it being mocked would be good. 😄