"ChangeSet" API is error prone in multi-threaded contexts

[LLFourn]

Non-monotone changes (i.e. SparseChain changesets) must be generated and applied in order. If in between generating a changeset and applying it you apply another changeset that changeset must be thrown away. This is especially likely to happen if you need to do some IO after generating a changeset before you apply it and so you drop your mutex lock on the chain allowing another thread to make modifications while you wait for the result of your IO.

We current have baked in this mistake into our electrum example and the idea the "inflate_changeset" API on chaingraph. There we first generate a sparse chain changeset, we do some IO to get the full transactions (and drop the lock), and then inflate the changeset and apply it. What if another thread makes contradictory changes to the blockchain during this time i.e. there is a new block and the blocks and transactions have moved position since then? It breaks. This particular approach can be changed by instead of keeping a changeset over the IO boundry we inflate to a full ChainGraph and apply that -- this will enforce consistency.

We may wan to come up with a more general solution that makes it impossible to make this mistake:

1. Changesets should commit to previous changesets to avoid applying out of order #70 would do this but we could also,
2. Do not allow applying changesets API but rather stage them inside the objects themselves and commit them when appropriate to return the changeset which can be persisted.

This does not need to be fixed for first release.