Modernize and cleanup locking

Allow serial to be returned as late as tpc_finish

This makes possible for storage to allocate serial inside tpc_finish,
removing the requirement to serialise 2PC's second phase phase (tpc_vote
to tpc_finish/tpc_abort).
Co-Authored-By: Julien Muchembled <jm@nexedi.com>

This also reverts commit 4905bb85:
sleeping 1ms is enough.

- Move from try/finally to with

- Start phasing out the _lock_acquire/_lock_release shortcuts

- Replace simpler @locked decorators with with statements.

  I've come to prefer the with style tp the decorator style.
  I think it's clearer.

  (The decorators preceeded the Python with statement.)

  I left the decorators in cases where they were used with preconditions.

This is mostly pretty mechanical, although it got a little delicate in
places and ... tests

Move load to imvcc storage

Which led to me to call site that needed it to be removed.

And finally get rid of the version argument.

Also drop the version argument from IStorageWrapper.invalidate.

Stop calling load

Load calls load before

Conflicts:
	src/ZODB/utils.py

Because the error semantics of load and loadBefore differ enough that
changing calls isn't trivial.  Plus the helper serves as a handy
transitional load implementation.

Simplify MVCC by determining transaction start time using lastTransac…

Obviously the tests don't call this. This is probably because they use
DB.storage (rather than DB._mvcc_storage).

Conflicts:
	setup.py
	src/ZODB/tests/synchronizers.txt

Abstract ZODB's MVCC implementation into a storage adapter.

- Lower-weight instances when not using RelStorage

- Still works with new RelStorage instances.

- Doesn't make me include loadBefore in MVCCAdapterInstance objects. :)

Hopefully, this will make linters happier.

I'd started to edit it to remove implementation details and got
distracted, leaving it half edited.

This allows us to not worry about method overriding and avoids
creating wrapper methods (some of which were broken).

My guess is that performance is a wash.  We take a tiny hit on the
first access, because ``__getattr__``, but we don't have to copy
methods we never use.

It wasn't really needed or used and caused a stupid ZEO test failure.

I do think the storage API needs to be thought about and refined at
some point, but now, it's not a priority.

I'd lazily assumed that undo and tpc_vote has the same sorts of
returns, but they don't.  When tpc_vote isn't None (ZEO), the retrun
is a list of oid and serial pairs.

To prevent modifying the set of instances while iterating over them.