mdb_cursor_put(): Use free sub-page space added by a
previous put(existing DUPSORT item), if any.
Tweak NUMKEYS code to avoid a new signedness warning.

It was inaccurate and misplaced. Verify it in mdb.c instead.

I.e. d69d2ce2, it left out some changes.

It and the MDB_msg_func can now return >= 0 for success.
Always return any MDB_msg_func() error result.

Use a sub-DB for DUPSORT item #1/#2 per key if needed: Not a sub-
page too big for a node, nor an overflow page (which not all DUPSORT
code checks for). Move "insert" code, to avoid non-loop goto upwards.

(This is the commit which needs the change to xdata.mv_size in
commit 9d6e4a91 "page sizes".)

Handle keys stored by a liblmdb with a bigger MDB_MAXKEYSIZE.
mdb_get/mdb_del(absent key bigger than our MDB_MAXKEYSIZE)
now return MDB_NOTFOUND instead of MDB_BAD_VALSIZE.

Change me_nodemax to not count the mp_ptrs[] entry. That's mostly
how it was used. Compare node sizes ">" me_nodemax instead of ">=".
The ">=" was a workaround for confusing sizes with and without the
mp_ptrs[] entry, but broke for nodes with size (old me_nodemax-1).

Explicitly make me_nodemax even. An odd value could break the
comparisons. It was even anyway because MDB_MINKEYS == 2.

Do not give uneven sizes to the sub-page and the node it lives in.

Don't set dkey.mv_size if mdb won't clear it before next iteration.

Drop unused computed/assigned values.
mdb_page_alloc(): Make "mop_len>0 implies mop!=NULL" provable.

Support MDB_NOSUBDIR. Catch more errors.  Write messages
to stderr, not stdout which the data too may be piped to.

On near match, return the matched data.

Measured perf cost of meminit was 7% worst case on an Intel Core2
system. YMMV.

Silence warnings, fix broken MinGW64 definitions, etc.

Init malloc'd memory by default, use #MDB_NOMEMINIT to disable.
Bump version to 0.9.10

Initialize unused portions of DB pages, for security or to silence
checkers like Purify and Valgrind. Like BDB configure --enable-umrw.

A node may need room for [key, subpage with 2 items] = 3 keys.

Use DB page size = min(32k, OS pagesize).  Previous limit was 8k
(MDB_MINKEYS*MDB_PAGESIZE).

Handle DB pagesize < OS pagesize. That's an I/O pessimization,
but transactions remain atomic: Only writing the MDB_meta must
be atomic, and it fits in one OS page.

Don't truncate desired subpage size: Asssign it to a size_t
(mv_size), not an uint16_t (mp_upper).

Skip cursors that aren't as deep as current cursor

Tweak split_indx check, go one slot further.

In d8eccb35