Fix black captures from servers announcing DesktopSize before pixel data (#90)

[sibson]

Fixes #90 — black screenshots against TightVNC on Windows, reported 2017–2025.

Mechanism (full analysis in the issue comment): TightVNC answers the first non-incremental FramebufferUpdateRequest with an update containing only the DesktopSize pseudo-rectangle. updateDesktopSize replaces self.screen with a black canvas, _doConnection fired commitUpdate unconditionally once the rectangle count hit zero, the pending capture completed, and the CLI saved the black canvas and exited — before the server's pixel frame (sent moments later) arrived.

The fix, two small parts:

• rfb.py: the DesktopSize rect is removed from rectanglePos before decoding, exactly as the QEMU key-event pseudo-rect already is ("undo append as this is no real update") — pseudo-rectangles carrying no pixels no longer count as update content.
• client.py: commitUpdate with a pending deferred and no content rectangles re-issues a non-incremental framebufferUpdateRequest and keeps waiting, instead of completing the refresh with nothing on the canvas.

A server that answered every non-incremental request with a content-free update could keep the client requesting until its timeout; real servers send content, and the per-client timeout still bounds the wait either way.

Regression tests live in their permanent home (tests/unit/test_client.py): a desktop-size-only update must not complete a refresh (and must re-request), and the refresh completes normally once pixel data arrives. Full unit suite green, flake8 clean.

End-to-end validation: the scripted server replaying TightVNC's logged behaviour (full repro kit in the issue thread) flips from a black capture (extrema ((0,0),(0,0),(0,0)), server logs could not send pixel data: CLIENT ALREADY DISCONNECTED) to the correct test pattern — after the DesktopSize rect the client now re-requests and receives the RAW frame before completing:

INFO:twisted:x=0 y=0 w=1024 h=640 <Encoding.PSEUDO_DESKTOP_SIZE: -223>
INFO:twisted:x=0 y=0 w=1024 h=640 <Encoding.RAW: 0>
INFO:twisted:Stopping factory <VNCDoCLIFactory ...>

Supersedes #342 (the test-only repro PR): this PR carries that test in final form, renamed for the behaviour it checks per the repo convention. When this merges, close #342 unmerged — its expectedFailure test would report an unexpected success on a fixed tree.

[sibson]

should commitUpdate even get called if we are properly deleting the pseudo rectangle