Per-column collation

Here is the next patch in this epic series. [0]

I have addressed most of the issues pointed out in previous reviews and
removed all major outstanding problems that were marked in the code. So
it might just almost really work.

The documentation now also covers everything that's interesting, so
newcomers can start with that.

For those who have previously reviewed this, two major changes:

* The locales to be loaded are now computed by initdb, no longer during
the build process.

* The regression test file has been removed from the main test set. To
run it, use

make check MULTIBYTE=UTF8 EXTRA_TESTS=collate

Stuff that still cannot be expected to work:

* no CREATE COLLATION yet, maybe later

* no support for regular expression searches

* not text search support

These would not be release blockers, I think.

[0] http://archives.postgresql.org/message-id/1284583568.4696.20.camel@vanquo.pezone.net