Improve spans when splitting multi-char operator tokens for proc macros.
When a two-char (or three-char) operator token is split into single-char operator tokens before being passed to a proc macro, the single-char tokens are given the original span of length two (or three). This PR gives them more accurate spans.