Summary:
Just skipping over a BOM, but leaving it in the Stringbuffer, is not
sufficient. The Lexer calls prevChar when a regular expression starts
with '^' (which is a shorthand for '\n^'). It would never match on the
first line, since instead of '\n', prevChar would still return '\xfeff'.

Test Plan: validate

Reviewers: austin, ezyang

Reviewed By: austin, ezyang

Subscribers: simonmar, ezyang, carter

Differential Revision: https://phabricator.haskell.org/D176

GHC Trac Issues: #6016