storage/engine/rocksdb: set optimize_filters_for_hits

[petermattis]

Do not create bloom filters for the last level (i.e. the largest level
which contains data in the LSM store). Setting this option reduces the
size of the bloom filters by 10x. This is significant given that bloom
filters require 1.25 bytes (10 bites) per key which can translate into
gigabytes of memory given typical key and value sizes. The downside is
that bloom filters will only be usable on the higher levels, but that
seems acceptable. We typically see read amplification of 5-6x on
clusters (i.e. there are 5-6 levels of sstables) which means we'll
achieve 80-90% of the benefit of having bloom filters on every level for
only 10% of the memory cost.

There is no significant impact on block_writer performance tests. The
most affected benchmark is:

name                             old time/op   new time/op   delta
MVCCGet1Version8Bytes_RocksDB-8   47.7µs ± 1%   53.3µs ± 2%  +11.69%  (p=0.000 n=10+10)

But this delta is a little spurious because the "old" data contained
4 sstables while the new contains 5.

Fixes #10050

---

This change is

[petermattis]

Cc @cockroachdb/stability

[tbg]

LGTM