Most of this is propagating the const-ness through the myriad write
functions.

Since the underlying kmem allocator uses the first sizeof(struct run)
bytes as metadata for freed runs, the zero pages put into the zero
page allocator's kmem were coming back with the first few bytes
non-zero.  Fix this by explicitly zero'ing the run header of pages
retrieved from kmem.

usertests now passes the bss test.  It still fails twofiles and
forktest.

Maybe we want vmnode to just take a byte size...

Previously, we mapped it as one node that spanned all of the pages,
but had a smaller size.  This causes a panic when we fault on pages
beyond the size because vmnode::loadpg calls readi with a too-large
offset.  We could "fix" loadpg, but what we have now is actually
POSIX-ish behavior (which says faults beyond the end of the mapped
object should signal SIGBUS), so instead we fix the ELF loader to
create two vmnodes in this situation.

This gets usertests running again for me (it now fails at the BSS
test, but at least it starts).

We insert dummy empty elements into crange and leak vmas in radix.

Otherwise a const markptr isn't all that useful.

Slightly easier to read.

I have a sucky assembler apparently

I have a sucky assembler apparently

The gcc docs say use "memory" if "your assembler instructions access
memory in an unpredictable fashion".  I don't know what that means,
but linux adds "memory" for similar inlines.  The gcc docs say use
"cc" if "assembler instruction can alter the condition code
register".  I know what that means, but linux doesn't use it.