Make sure that loadBefore actually uses ._conflict_cache data

TemporaryStorage uses

    ._index   {} oid -> serial
    ._opickle {} oid -> data

as data store for load. However for loadBefore

    ._conflict_cache  {} (oid,serial) -> (data, time)

is reused as data store.

That would be ok if that place would be treated as data store, but given
its primary purpose - as its name suggests - was originally to be a
cache to resolve conflicts, it is "logical" that entries in this cache
are garbage-collected when entry age becomes > gc threshold.

Only now there is a problem: if an object is committed once, and time
passes, corresponding entry in ._conflict_cache will be removed. And this
would manifest itself as

    - load(oid)  -> gives latest data for the object (obtained via ._index and ._opickle)

but

    - loadBefore(oid, @head)  -> gives POSKeyError

-> Fix it by always preserving latest object revision in
._conflict_cache from being removed on GC.

The fix is important for systems that use ZODB5, or ZODB4-wc2[1] because
there ZODB.Connection switched from primarily using load to exclusively
using loadBefore.

/cc @icemac @mauritsvanrees @mgedmin @d-maurer @dwt @hannosch

Fixes: https://github.com/zopefoundation/tempstorage/issues/8
[1] ZODB@8e7eab33

This prevents a KeyError: '\x00\x00\x00\x00\x00\x00\x00\x00'

I guess there were no tests of GC.

That was made effective by recent changes in ZODB that relied more on
loadBefore for MVCC.

Oddly the change makes this loadBefore match the version it was copied
from, which github says was last changed 13 years ago. (gulp, on
multiple levels).

See: http://blog.travis-ci.com/2014-12-17-faster-builds-with-container-based-infrastructure/

Downloaded from http://downloads.buildout.org/2/bootstrap.py

Python 3 compatibility

Make the local tests use "normal" unittest conventions.