reduced overhead of MDB_RESERVE

[nolanderc]

See also my changes here https://github.com/nolanderc/heed/tree/reserved-space.

• use #[inline] on ReservedSpace to enable inlining across crates. It is also possible to use lto = "thin" or lto = true under [profile.release].
• Made it possible to write into &[MaybeUninit<u8>] without UB using UninitWriter.

Seems std::io::Write for &mut [u8] is slower than expected. In my experiments, moving the check for buf.len() > space.remaining() as early as possible in the write function improved performance drastically (2x faster).

Test results on my laptop:

$ ./run.sh
+ ./target/release/heed-perfs-put-reserved classic-codec
  19.61s [insert:    8.04s, commit:   11.57s]
+ ./target/release/heed-perfs-put-reserved put-reserved
  19.37s [insert:    7.74s, commit:   11.62s]
+ ./target/release/heed-perfs-put-reserved put-reserved-alloc
  20.05s [insert:    8.83s, commit:   11.22s]
+ ./target/release/heed-perfs-put-reserved put-reserved-uninit
  18.57s [insert:    6.66s, commit:   11.91s]
+ ./target/release/heed-perfs-put-reserved put-reserved-uninit-fill-zeroes
  19.00s [insert:    7.10s, commit:   11.90s]
+ ./target/release/heed-perfs-put-reserved put-reserved-uninit-into-slice
  26.32s [insert:   15.40s, commit:   10.92s]

There's not a huge improvement over just pre-allocating a Vec, but yielded a 1.6s reduction in insertion time at its best. But this is a small benchmark where all data is the same size, which is also done on my laptop, so YMMV.