Declare the fdv1-fallback capability and accept the new top-level
dataSystem.fdv1Fallback config object, wiring it directly to the SDK's
FDv1 Fallback Synchronizer. Drop the heuristic that previously inferred
the FDv1 fallback from the last polling synchronizer entry -- the
FDv1 Fallback Synchronizer is a distinct configuration concern from
the FDv2 Primary/Fallback chain.

Also bump the test harness pin to v3.0.0-alpha.6 so the new
fdv1-fallback test suite runs in CI.

Per the FDv1 Fallback Directive (Data System spec 1.6.3), the Primary
Synchronizer must be terminated when the directive engages. The
streaming source called ``SSEClient.close()`` to do this, but on
Python 3.10 that wasn't enough: ``SSEClient.close()`` ultimately
invokes ``urllib3.HTTPResponse.shutdown()`` (which only performs a
half-close on the local read side) plus ``release_conn()`` (which
returns the live connection to the pool). The TCP socket stayed open
until the response was garbage-collected, which the contract test
harness flagged on the linux/3.10 matrix entry as
"SDK did not close the FDv2 streaming connection after the directive
— the Primary Synchronizer must be stopped when Directed Fallback
engages." Newer Python versions GC the response promptly enough to
hide the leak.

Track the urllib3 PoolManager alongside the SSEClient and call
``HTTPConnectionPool.close()`` on each pool when the synchronizer
stops. ``HTTPConnectionPool.close()`` drains the connection queue and
closes each socket, sending the FIN the server is waiting on.
``PoolManager.clear()`` alone won't do this -- it drops the dict of
pools without closing the connections inside them. The fix is
deterministic (no sleeps) and runs both at the natural end of
``sync()`` and in ``stop()`` so it's robust to whichever thread
arrives first.

Two follow-up cleanups from review of the FDv1 Fallback Directive work:

- Drop the test-only shim in StreamingDataSource.sync() that accepted
  either a bare SSEClient or a (client, pool) tuple from the injected
  builder. The SseClientBuilder type alias already declares the tuple
  return; annotating the field as SseClientBuilder lets pyright align,
  and the lone test fixture (list_sse_client) now returns
  (client, None) like the production create_sse_client does.
- Hoist basis.change_set.selector.is_defined() into a local in
  FDv2._run_initializers() so it isn't computed twice on the
  success-with-fallback path. Also fixes a typo in the adjacent
  comment ("if an only if" -> "if and only if").