Improve thread-death error reporting in tests.

Reworked the way some of the ZEO tests use threads, so that unittest is
more likely to notice the real cause of a failure (which usually occurs in
a thread), and less likely to latch on to spurious problems resulting from
the real failure.

Detail:  The TestThread class got the (unittest) test case as an argument,
and remembered it.  If the run() method saw the thread die with an
exception, it called testcase.fail() with the formatted exception.  But
this can't work as apparently intended:  unittest's fail() raises an
AssertionError then, and the piece of unittest running the test expects
to catch that exception and append the message to the result object's
list of failures.  But the piece of unittest running the test isn't on
*this* thread's call stack!  It's on the main thread's call stack.
So unittest simply raised an exception that wasn't noticed by anything,
and the thread died then, leaving a traceback somewhere in the middle
of the output.  unittest didn't know anything about that, and sometimes
thought the test passed despite the thread failure.

Alas, when a thread dies like this it's also got a decent chance of
leaving things in a state where the test's final checking code can't
succeed, but the checking code doesn't know the thread died either,
and spurious unittest failures and errors could get reported then.

Reworked things so that only the main thread ever tries to tell
unittest that a test failed.  A TestThread remembers an exception-
death now, but doesn't report it until the main thread tries to join
it.  This way unittest knows the test failed; records and reports
the true cause of failure; and, because tests generally join all their
threads before doing their final checking, unittest stops the test
before final checking if a thread death occurred, and so there's
less chance of seeing reports of spurious errors and failures.

Write backups to a temp file first; flush and fsync when done;
only then rename; in case of a crash, no relevant damaged file
will be left behind then.  And on Windows, Python's os.fsync()
is necessary to convince the OS to write anything to disk.

Repaired massive lies I introduced about wether fsrefs requires
application code to be available.  Noted that fsrefs also partly
verifies the sanity of the .fs.index file.

Repaired a bug wherein spurious error msgs could be produced after
reporting a problem with an unloadable object (discovered by eyeball,
while staring at the code to figure out what it actually does).

Vastly expanded the module docstring, with a slimmed-down version of
the new fsrefs docs on the ZODB Wiki.

Connection objects were typically immortal because the threaded
transaction manager kept them in an ever-growing lists.  Reworked the
transaction manager internals to use a simple implementation of weak sets
instead.  This plugs all leaks in the test program attached to the
collector report (which was leaking about 100KB/sec on my box).

Forward-porting from Zope 2.7.

The checkFullTimeStamp() test was sensitive to unique mktime() behavior
on FreeBSD.  See:

http://lists.freebsd.org/pipermail/freebsd-standards/2003-November/000268.html 

The purpose of this test is to exercise ZODB's TimeStamp object, so got
rid of dependence on platform mktime() and time.timezone quirks --
TimeStamp works in GMT, so how mktime() treats tm_isdst should be
irrelevant in all TimeStamp tests.

Also added a comment about the highly non-obvious numeric characteristics
of TimeStamp's treatment of seconds (round-tripping is surprisingly
inaccurate, but for a real reason).

If the system clock moved back in time more than 30 days between
a FileStorage close and (re)open, new txn ids could be smaller
than txn ids already in the database.  Plugged that hole.

New test checkRestoreWithMultipleObjectsInUndoRedo from ZODB 3.2.2.
3.3 didn't have the bug, and this confirms it.

move the release metadata for zpkg here from the CVS repository;
this was in Releases/ZODB3/ on cvs.zope.org

worker thread creates could cause mysterious exceptions in
*later* tests (when running all the tests, they appeared
to come out of checkDetail, which runs right after
checkMinimizeTerminates).  Changed the CantRidOfMe class
so that instances of the class *can* be gotten rid of,
keying off a new module bool.  The checkMinimizeTerminates
worker thread uses this to defang the objects after the
body of the test has completed.

In replace(), print feedback to stderr, since fileinput
hijacks stdout.

release, tried to make this script much more robust.  Use
r-strings for regexps.  Junk unnecessary code that a Unixhead
probably thought was needed for Windows (replacing forward
slashes in filepaths -- Windows accepts those too).  Made
the search criteria for NEWS.txt much looser (this is the
one we seem to blow most often).  Changed replace() in
several ways:  only replace the first occurrence of the
pattern (we never need more than that, and looser searches
are dangerous when replacing all occurrences); provide
feedback on everything done; if the pattern isn't found
at all, print a noisy msg saying so.  Completed the list
of modified files in the module docstring.

SAX stuff (that all comes with the minimum Python needed
to run ZODB now).

God knows it will get out of synch someday, but continuing to call
it version 0.3a2 is crazy.

Call it beta 1.  Recommend Python 2.3.4.  Retract the promise that ZEO incompatibilities will be fixed in
beta 1.

to the dead StandaloneZODB project at SourceForge.

cacheMeanDeal, and cacheMeanDeac methods from the DB class.  They
were untested, undocumented, unused, and did nothing of use (the
first always returned an empty tuple, the rest always returned None).