The problem is essentially that if we try to parse a token tree using a
CodeMap different from the one the tree was originally parsed with,
spans become nonsense. Since CodeMaps can't be cloned, we're basically
forced to use the original ParseSess for additional parsing.
Ideally, rustfmt would be a bit more clever and figure out how to parse
macro arguments based on the definition of the macro itself, rather than
just guessing that a particular token sequence looks like an expression,
but this is good enough for now.
Fixes#538.
This removes usage of:
* PathExt
* split_last
* split_last_mut
* catch_panic
The catch_panic one was a little tricky as the ident interner needed to be
cloned across threads (a little unsafely), but it should otherwise be good to
go.
