In preparation for actually making page reads async, split page reading
into two parts: I/O submission with `begin_read_btree_page()` and
parsing with `finish_read_btree_page()`. Also wire up completion to
perform the read parsing.
Next step is to move the Completion.complete() call somewhere in the I/O
layer and make sure we return from the VM correctly if page is under
I/O.
We need to be able to allocate a new page and insert it into a page
cache without contents for asynchronoous I/O. Let's do that by making
`contents` optional in Page. (We perhaps ought to rename it to
`inner`...)
We want to submit I/O asynchronously with io_uring, which means we can
have multiple concurrent lookups to the page that is under I/O.
Therefore, let's add three page flags: up-to-date, locked, and error, so
that the higher level layers can just look up a page, but wait if
there's I/O happening.
