Issue Description:

We mostly think of trivia as things that could be dropped without changing semantics, but some are crucial. For example, the space in “let x=1” prevents tokenization as “letx, =, 1.” Also, the newlines between statements, where there are no semicolons, are required. A formatter/pretty-printer that normalizes whitespace needs to be able to find-and-preserve or ensure-it-recreates nontrivial whitespace.

The current libSyntax API makes that especially difficult when multiple identifiers and/or keywords appear together, because the rules about what characters can be part of an identifier are complex. One could special-case all the known token combinations, but when the language grows a new token these rules are very likely to break down until the formatter is updated and rebuilt with a new libSyntax. It should be possible to handle this in an efficient way that is very likely to be resilient to future language changes.