trie: avoid rlp-encoding when trie hash is already available

[holiman]

I'm not sure this is correct, of if I'm missing some quirky edgecase, but what this PR does, is attempt to improve on the scenario where we first Hash the trie and later Commit it.

That typically happens on every block, since we call IntermediateStateRoot after applying the block rewards, and that does Hash internally. We need to hash it to get the state root, and we need that in order to validate against the header.
If everything checks out, we then Commit. Now, the Commit is clever enough not to re-hash everything since it already has the hashes cached for all nodes. However, we do rlp-encode everything the second time around, which this PR skips if we have the hash cached.

Turns out that saves us 20% processing time:

[user@work trie]$ benchstat before.txt after.txt 
name               old time/op    new time/op    delta
CommitAfterHash-6    5.08µs ±12%    3.99µs ±21%  -21.33%  (p=0.000 n=20+18)

name               old alloc/op   new alloc/op   delta
CommitAfterHash-6    1.79kB ± 1%    1.65kB ± 1%   -7.67%  (p=0.000 n=18+17)

name               old allocs/op  new allocs/op  delta
CommitAfterHash-6      11.0 ± 0%      10.0 ± 0%   -9.09%  (p=0.000 n=20+20)

[holiman]

this doesn't work, the code below assumes h.tmp contains the rlp-encoding of the node :(

[holiman]

But actually, the db.insert only cares about the size (for tracking), not the actual data