This avoids unnecessary rewrites of pages that do not change.
(Restructuring for upcoming mdb_page_spill work.)

TODO: Rename C_UNTRACK to C_TRACKED.  Omitted now for readability.
The current name is because it's lazy: not always set when tracked.

(Restructuring for upcoming mdb_page_spill work.)
mdb_freelist_save() can't just Get() the destination, since
mdb_page_spill() may have put the destination in the read-only map.

TODO: Can this new put() modify the freelist, which would break it? The
final iteration's put() can shorten the node, the rest uses MDB_CURRENT.

We could set P_KEEP on dirty freeDB leaves and ovpages, since they are
all about to be modified.  But the code in this commit must stay anyway,
if mdb should support dropping a 256G DB. I.e. too big for dirty_list.

Commit d6d2638a broke read
on zero-length files.

Nor uninited cursors' subcursors' page pointers.

When collapsing root, must also move cursor index down,
not just the page pointer.

Also in mtest, break from NEXT loops on error, otherwise it just
prints the previous key/data again, which looks confusing.

If mdb_page_touch() sees a page in txn's dirty_list, that
is the page version txn's cursors should have.  Fail if
the user may be seeing and depending on another version.

Restore mc_flags and xcursors, they were tracked but not merged.
Simplify: Track parent txn's original cursors after backing them
up, instead of tracking copies and merging them back at commit.

Terminate loop on intermediate failures, return count of written items,
document usage.

Page leak, mdb_page_alloc(). On error, don't shorten me_pghead.
Memleak, mdb_ovpage_free(). Free page or keep it in dirty_list.
Bad MIDL, mdb_midl_need(). Fix midl[-1] (allocated size).

Catch I/O errors. Do nothing between OS call failure and ErrCode().
Do not use errno after non-OS-errors like write() >= 0, which could
give a failure return of success (errno 0) or some irrelevant error
code.  Drop seek calls, use pwrite/pread/Windows OVERLAPPED offset.

Don't put a 64-bit filesize in a 32-bit int before shifting
down.  Always pass &sizehi to SetFilePointer->maxsize, so
sizelo not is treated a signed distance.  Hide unused vars
when _WIN32.  Reinitialize OVERLAPPED before reuse.

Grow midls earlier in order to catch errors earlier.  Use
mdb_midl_need() instead of mdb_midl_grow(), then mdb_midl_xappend()
needs no error checks. Factor out mdb_midl_append_range().

mdb_page_malloc(): Take a txn arg instead of a cursor.

Split out mdb_dpage_free(), mdb_page_flush() and clean up.

Merge if() branches. Restore retry=500 when MDB_PARANOID, for clarity.

MDB_env.me_pghead: Don't free it when empty. mdb_ovpage_free()
needs it, but cannot allocate it.
mdb_midl_alloc(): Fill in length=0.
mdb_page_alloc(): Also Skip freeDB if txnid<3, instead of <4,
and consistently DPRINTF consumed IDLs.

Committing a nested txn lost the MDB_TXN_DIRTY flag
in the parent, unless the child had set it too.

When copying, round up/down to aligned sizes. Skip the unused portion,
this was not done when touching a page dirty in the parent txn.
No other change in behavior.

Simplify mdb_page_touch(), including: Drop test m3==mc, the condition
is caught below. Don't "modify" the parent's pgno into the same pgno,
when a nested txn copies a parent's page into its freelist.

The tracking code should not change the current cursor.
It did when that was a C_SUB cursor, which should not be
checked against the tracked cursors but their xcursors.

However, do not bother to skip the tracking code for the
current cursor when it would not change that cursor anyway.