Convert v1 to v2 range-del blocks on the fly

[petermattis]

RocksDB generates range-del blocks that contain range tombstones that are
not fragmented and sorted. Pebble generates range-del blocks containing
tombstones that are fragmented and sorted. The fragmented and sorted
range-del blocks are suitable for serving from directly, while the
unfragmented/unsorted blocks are not. We now transform RocksDB format
range-del blocks into Pebble format blocks on the fly.

The transformation is done via a hook that allows passing a block through a
function after it has been read from disk but before it is added to the
cache.

Fixes #68

Provide key stability across blockIter positioning calls

Provide stability for the keys returned from a blockIter across positioning
calls (Seek*, First, Last, Prev, Next) when the block has a restart
interval of 1. This is achieved by noticing when a key does not share any
bytes with the previous key and changing the returned key to point directly
into the block instead of the key buffer. This behavior was already being
expected by rangedel.Truncate and by forthcoming code to transform v1
range-del blocks to v2 format on the fly.

[petermattis]

This change is 

[petermattis]

I need to add testing for the v1 -> v2 transformation, which is why this is still marked [WIP]. The first commit is ready to go and something I just stumbled upon. I'm not entirely settled on that approach. Perhaps it would be better for rangedel.Truncate and this now range-del block transform to copy the keys returned during block iteration.

[petermattis]

Tests have been added. I did this by adding a private Writer.rangeDelV1Format flag that causes an sstable writer to allow unfragmented and unsorted tombstones to be added to the sstable. I think this PR is ready to go.

[Ryanfsdf]

Reviewed 2 of 2 files at r1, 5 of 5 files at r2, 2 of 2 files at r3.
Reviewable status: complete! all files reviewed, all discussions resolved (waiting on @ajkr)

[petermattis]

TFTR!