Such that we don't break debug build all the time.

Such that we don't break debug build all the time.

Such that we don't break debug build all the time.

Such that we don't break debug build all the time.

Caught by John Daggett.

Also remove wrong ASSERT_SIZE that would trigger only when not using
flexible arrays (which was the case on win32, but not with gcc).

The major-version check is now handled by sanitize.  If major
doesn't match, we reject and fall back to the Null object.

Before, as soon as we needed to make an edit, we tried to make the blob
writable inplace.  That grows code unnecessarily though.  We can simply
fail, make writable, then start again.  That's indeed what the fallback
was doing anyway.

Two things:

1. Allow nested Extension lookups.  The offset is always positive, so
it can't loop circularly.

2. Move the check for all Extension subtables having the same lookup
type to the correct place.  Before it wasn't really working.

We handle major-version differences via get_for_data(), so sanitize
should never see a major version mismatch.