In-memory trie operations

[rphmeier]

This is an optimization PR introducing an alternative implementation of TrieDBMut.
While the old TrieDBMut recalculates all affected node RLPs and commits to the DB at the end of each insert or remove operation, the new one does only when it is dropped or root or commit is explicitly called.

[arkpar]

returning ElasticArray1024 here would save lots of memory allocations
stream.drain()

[rphmeier]

So this implementation has a flaw where it never deletes nodes from the DB and may re-insert nodes that already exist in it.

[rphmeier]

gives ~5% speedup importing the first 250k blocks for me

[coveralls]

Coverage decreased (-0.06%) to 76.424% when pulling e64d6a7 on rphmeier:insertion_trie into 4226c0f on ethcore:master.

[gavofyork]

coverage decreased?

[gavofyork]

} else { line is a bit wasteful for these trivial ternary-like conditionals. typically use match instead.

[gavofyork]

a few points where style could better follow the "house rules". otherwise looks like quality code. didn't do a deep inspection of the logic, which is likely unnecessary assuming all tests pass and unlikely to uncover any deep-seated bugs.

probably good to get another pair of eyes over it though.

[coveralls]

Coverage decreased (-0.02%) to 76.461% when pulling a9de714 on rphmeier:insertion_trie into 4226c0f on ethcore:master.

[rphmeier]

coverage decreased as a percentage because the code itself is larger. i have also added an insert_empty test because this is where consensus failed previously. this reduced the amount that coverage decreased by.

[coveralls]

Changes Unknown when pulling 9c36b5f on rphmeier:insertion_trie into * on ethcore:master*.

[rphmeier]

travis failed on an unrelated network test, but not before all trie json tests and util tests passed.