fix: compaction support pick by max_seq (#1041)

## Rationale
Part of #987.

Current implementation will compact by file size, max_seq is not
considered, this may cause data corruption
in corner case, eg:
- sst1, max_seq:10, PK1=10
- sst2, max_seq:11, PK1=9
- sst3, max_seq:12, no PK1

If compact pick sst1 and sst3, and output sst4, its max_seq will be 12,
now PK1 exists in two files:
- sst2, max_seq:11, PK1=9
- sst4, max_seq:12, PK1=10

That's to say, PK1's value is 10 now, which is wrong value(9 is right).

## Detailed Changes
When do compaction, first sort sst by max_seq desc, then only pick
adjacent ssts, the original issue is fixed in this way.
At the same time picked ssts are ensured to meet other requirements such
as `min_threshold`, `max_threshold`, `max_input_size`.

## Test Plan
UT and manually.