and fix the implementation to read the configuration correctly!  thanks, test.

This test is bogus, though, because the Tahoe-LAFS web API test double
misbehaves for directory entry requests.

The capability encoding misbehavior breaks some of our unit tests which are,
supposedly, using the API correctly.

hooray, more than a one-off fix

Previously we had it searching [1, inf] to discover the boundary at 2 ** 63.
Now we bias its distribution exponentially and hopefully find the very large
boundary sooner.

Also, remove a data() usage as long as I'm here.

If there are not enough reachable storage servers, try again.  Usually the
client node will manage to connect to some more eventually.

this will come in handy for building strategies for testing SQL interactions

Before it just ended up with DeferredResource which proves nothing at all.
Now it requires a sufficiently non-error-y status code to be happy.

Add a GET variation of the endpoint to poll progress and discover completion.

Also factor out the details since we would otherwise have four instances of
them now.

Maybe it always should be?  But configuring only PYTHONPATH is strictly easier
than configuring PATH and PYTHONPATH so I guess we'll let people do that.

By the time this code runs we know PYTHONPATH must be configured correctly
since we import stuff from allmydata elsewhere.

This helps with debugging failures.
Also the directory deletion implementation is broken on Windows anyway and
sometimes fails with errors like:

```
PermissionError: [WinError 32] The process cannot access the file because it is being used by another process: 'C:\\Users\\RUNNER~1\\AppData\\Local\\Temp\\tmpc8za73y6\\private\\privatestorageio-zkapauthz-v1.sqlite3'
```

failing because somebody leaks a Client into the reactor