Splitting a word into MSB/LSB pairs happens often and is worth, I
think, native words. Also, I'm going to need it in the upcoming
commits.

I wanted to make CVM's forth use the Grid subsystem, but doing so
would break the stage binary. Hence, this split.

I've noticed that under certain conditions, such as a stack overflow,
I would segfault, something I though was impossible because my memory
size is 0x10000 and all my integer variables are uint16_t.

After having re-read my C handbook, it seems I wasn't sufficiently
knowledgeable about type conversion rules.

There are many situations where it can be useful. Worth it.

I'm pretty happy about how lightweight the implementation turns
out to be.

Adding a delay such as the 20ms one we have in AVR programmer's
initialization routine is tricky without a word like TICKS.

This implementation is highly inaccurate, but more accurate and
reliable than a "ballpark" DO..LOOP...

This should help spot bugs due to bad initialization.

They were a bit terse.

The 1 byte limitation has been effective for a while now, but I
hadn't made the move yet, I wanted to see if the limitation would
cause me problems. It doesn't.

Doing this now slightly facilitates the IY->BC move in z80.

Bootstrapping: if you try to recreate the CVM binary from the
previous commit with this code, you'll have bootstrapping problems.
The first bootstrap will compile a binary with 2-bytes wide cells
but branching conditionals that yields 1-byte cells. That's bad.

I got around the issue by temporarily inserting a "397 399 LOADR"
instruction in cvm/xcomp.fs, right before the xcomp overrides. This
way, I force 1-byte cells everywhere on the first compiliation,
which then allows me to apply the logic change in cvm/vm.c and have
a properly running binary.

This will soon replace the libz80 based "forth" and
"stage" executables. This is much, much faster.