And fix bugs found by these tests...

Conflicts:
	unix.rs

First non-trivial change for a while... because this involves reading from the OS, we should cache as much as we can in memory. So, group membership checking is done immediately after reading a group name, as the group structure has already been populated.

Curiously enough, this turned out to be helpful today... Git can't
traverse symlinks, so I had to use hard links instead, and then I had to
work out which files had more than one link at a glance. If I hadn't
already had most of exa written, I'd be stuck using find!

This is behaviour cribbed from ls. Even though I've never personally used it, I'm sure someone has a use for it.

This requires the dir field to store a reference to the path that
created it, so that path can be used to join onto the filename so it can
be properly statted.

There doesn't seem to be an io::FileType entry for character special
devices or sockets, so these fall under the io::UnknownType entry. Which
also gets highlighted in yellow, for precisely that reason.

I checked strace and it was reading /etc/passwd every time it needed to
look up a user or group. Now it only does it once per.

Dir::readdir and File::from path now both return IoResults, rather than just calling fail! and exiting. This allows the program to continue after an error.

This left-aligns the final column, but the final column is the file name, so that's okay.

Fall back to strings. Also, add some unit tests.

- s/syntax/phase
- no more boxed array syntax
- getopts fail implements Show
- option#expect now requires &str

I'd much rather have this separate, as it has the definite potential to
balloon up to a huge size and end up making the other parts of the file
module hard to read. But on the other hand, it meant making a few more
methods public on file... readability over cleanliness, I guess.

This new object is then passed down to the File, so it can see what files are around it. This means it doesn't need to re-query the filesystem several times per file, instead using the in-memory copy.

Also, switch to using the built-in Path#with_extension method.

And an extra level of compiled files that don't get highlighted when their source files aren't present. I don't really like having two functions that do (almost) the same thing, but I have to due to the way colouring is done.

Compiled files with a source present (such as code.o when code.c is present) are now classified as temporary files. If the source isn't present, they're highlighted in a kind of drab colour (using the new 255-colour ability, yay!)

The code does do exists() checks on the filesystem when it could be possible to compare the files to the list of files we got from the call to readdir(), but it doesn't.

Add a Fixed(u8) constructor to Colour, which represents the 256 colours
that some terminals support. This means we can:

- stop using black bold to mean grey, which looks weird on terminals
  that haven't been set up to use it;
- support a *lot* more file type colours.

I'm a little suspicious of how much string allocation is being done in
colours.rs, but that's a problem for another time.

I think this is necessary because 'bytes' currently has no 'B' suffix, and it's
kind of hard to distinguish a long number from a suffix.

Currently there's only one numeric column, and that's the file size, so it gets
special treatment.

I was originally going to have a folder file size field be filled up with '-'s
as far as it could go, leaving it entirely up to the column how its field gets
formatted. But then I saw just one '-' working just fine, so I left it like
that. In the first try, columns could do anything they want when padding a
string (including changing the padding character or just changing it entirely),
but now there's no point.

- s/to_owned/to_string/g
- TotalEq/TotalOrd changes
- snake_case of identifiers