Fixes #6861.

`current` might not be a prefix of the (normal) completion candidate
`path` if `current` itself is non-normal. In case `current` is longer
than the candidate `path`, the code previously panicked. The tests have
been extended to trigger this panic.

The panic is avoided by using `strip_prefix()` instead of direct
slicing. However, in case `current` does not prefix `path`, the path
may be truncated too early. Additionally, fish discards any completion
candidates that do not start with `current`, even if the completion
script writes them to stdout.

Thus, this commit is the minimal change to avoid the panic, but it does
not yet handle completion of non-normal paths correctly.

Rewrites how the file-path completion is determined from a potentially
non-normal `current_prefix`. The basic idea is to normalize the
`current_prefix` first to obtain an actual prefix of the (normal)
`path`. The remainder can then be spliced onto the original
`current_prefix` to form a non-normalized completion path that is
prefixed by `current_prefix`, allowing the shell to accept it.

This requires a different API for the helper than what `dir_prefix_from`
provided. The latter assumed that when `None` was returned, `path` could
be completed as-is; a `Some` value indicated a partial completion of
`path` to the next directory. This is no longer sufficient as we need to
potentially return a different, non-normal path in either case.

Since we want to ignore the `mode` for directory completions, the
helper must also return the type of completion (partial/directory
or full/file). To avoid propagating this information, I instead
pass the `mode` into the helper and have it return finished
`CompletionCandidate`s. This also yields cleaner code on the caller
side.