When we don't have any registers to pass, it's convenient to be able
to share a common all-zero structure.

Yet another byte bites the dust...

Shave off another byte by using the not-frequently-used jmpw *%sp
instruction!

Comments are good...

Use the "return address points to data" trick to shave a few bytes off
the generation of error messages.

In case we hit the death loop, put HLT in it.  It's probably superfluous,
but it can't hurt, and we can always take it out to reclaim the space.

Arguable the death loop is superfluous at all, since int $0x18 should
never return.

Six bytes down the drain, but people do expect it.  If it becomes
an issue we can remove the period; it's better than the CR LF.  There
might be some way to efficiently issue . \r \n in common code, too.

We always re-read the last partition table on any kind of error, so
no need to prevent the code from clobbering it on failure.