Previously, when we read-faulted on a file-mapped COW page (which can
happen because we don't load anything in to a forked address space's
page tables), we would unconditionally readi to fill that page's
contents, even though that page had often been allocated and filled by
another process (possibly long ago).  This combines allocating and
loading a page into one atomic step that only happens if the page has
not already been loaded into the vmnode.

Coincidentally, this eliminates all unexpected sharing from dirbench.

So scale-amd64 can build lwIP again..

This have to do with implementation, not interface.  Removing these
unmasks the sharing on tlbflush_req, which is something we expected.
Luckily, it doesn't expose any other sharing.

Load indirect array on demand and flush only on iupdate.