We use a Latin1 generated parser with Alex, but we also parses Unicode
BOM, unbreakable spaces, etc. In recent Alex, the reported column isn't
expressed in Unicode chars anymore but in bytes/ASCII chars (probably
due to
https://github.com/haskell/alex/commit/ae525e34edf017544e8ef4457d7e57cf2081dcf9
but I haven't checked), which broke our tests (see
https://github.com/haskell/cabal/pull/8896).

To work around this we report indentation warnings at token start position,
instead of token end position (i.e. always 1). Otherwise position makes
no sense anymore for the user.