More performance optimizations

[arkpar]

• Moved DB commit out of the import lock. DB WAL should not hurt performance anymore on a multicore system
• Moved RLP compression to background as well
• Reduced last_hashes cloning
• Other minor tweaks

[coveralls]

Changes Unknown when pulling f723e57 on perf into * on master*.

[coveralls]

Changes Unknown when pulling 9b87657 on perf into * on master*.

[gavofyork]

does the DB write queue form a read cache, also? otherwise how is cohesion enforced while the write-queue is non-empty?

[arkpar]

It does, that's the whole point

[rphmeier]

seems like all of these Arc<LastHashes> could just be a &LastHashes and we would save a few atomic ops in a bottleneck.

[rphmeier]

and then since LastHashes appears to be just a Vec<H256>, why not just have the last hashes field be a &[H256]?

[arkpar]

Changing EnvInfo to hold a reference is much more involved. I'll leave it to future PR.

[rphmeier]

a few issues in the unsafe code introduced here and still a little more room for optimization, otherwise LGTM

[coveralls]

Coverage increased (+0.02%) to 86.605% when pulling 1613bcc on perf into deceb5f on master.