feat(swift): support teach() teach panel for Swift-backed robots

[petercorke]

Summary

robot.teach() only ever worked for PyPlot/PyPlot2 -- Swift's connector wrapper explicitly set supports_teach=False, so any hasgeometry=True robot (the default backend Swift resolves to, matching plot()) raised a clean TypeError. examples/teach_swift.py was RTB's own hand-rolled template for what a Swift teach panel looks like, but predates Swift 2.0's AssemblyHandle refactor and still drove the robot via the now-deprecated robot.q[j] = value direct-mutation style.

What's implemented

Swift._add_teach_panel() using the current idiomatic pattern (named sliders -> env.values -> one per-step handle.callback, see examples/panda_ik_sliders.py): one slider per joint (degrees for revolute, native units for prismatic -- teach_swift.py's always-degrees conversion was a bug, not carried forward), plus a live end-effector pose readout (6 compact Labels, matching PyPlot's own six fig.text() calls).

Dependency on companion swift PRs -- CI will be red until these merge

This needs swift-sim with:

• jhavl/swift#131 -- Label(compact=True), used by the pose readout
• jhavl/swift#132 -- graceful disconnect-during-run() handling, needed so closing the browser tab mid-teach ends the panel cleanly instead of an uncaught TimeoutError

RTB's CI already installs swift from git+https://github.com/jhavl/swift.git@future (not PyPI, see ci.yml), so this will resolve on its own once both merge -- no CI config change needed here.

Also fixed, both blocking without them

• teach()'s env.launch("Teach " + self.name, limits=limits) passed the name positionally, landing on Swift's real launch(realtime=..., ...) rather than the PyPlot launch(name=..., ...) it was written against -- raises ValueError immediately. Fixed to name= as a keyword.
• Swift's hold() (what teach()'s shared if block: env.hold() relies on to keep the panel open) only sleeps and polls for a disconnect -- it never calls step(), so nothing would ever process a dragged slider. _add_teach_panel() now runs its own env.run() loop instead when block=True, matching every other interactive Swift script's own step loop, and signals back to teach() so it skips the now-redundant env.hold() call. That call wasn't just wasteful to leave in -- it could hang indefinitely (headless mode's hold() never reports "disconnected" at all; even non-headless there's a race around close() and socket.USERS), which is what the two swift PRs above close off.

Other things worth knowing

• teach()'s docstring now documents that robot.q holds the final taught pose once teach() returns, true for every backend. PyPlot achieves this by mutating robot.q throughout its own session; Swift's _add_teach_panel() writes handle.q back once, at the point the session ends, since AssemblyHandle deliberately never mirrors it during the session itself (Robot/Shape "instance handle" redesign (animation-loop API) jhavl/swift#85). A one-time, deliberate exception to "stateless robot model" for this specific single-owner interactive session, not a general precedent -- the same tension PyPlot's own teach() already has (desiderata.md), just accepted here rather than solved.
• rtb.models.Panda().teach() with no explicit backend= now opens Swift by default (hasgeometry=True resolves there, matching plot()'s existing default) rather than always falling back to PyPlot, since Swift now actually supports it -- a real, intentional behaviour change, not incidental.

Test plan

• tests/test_backend_capabilities.py: TestSwiftCapabilities's supports_teach assertions flipped False -> True
• tests/test_BaseRobot.py: 3 new tests -- explicit backend="swift" smoke test, default-backend-resolves-to-swift regression test, and a test confirming robot.q is written back after the session ends (with Swift.run mocked to a no-op, since block=True's real loop only exits on a real disconnect)
• Full suite passes: 91 passed (test_BaseRobot.py + test_backend_capabilities.py)
• Manually verified end-to-end in a real browser: sliders drive the robot, pose readout updates live, all 7 Panda joint sliders visible without scrolling, closing the tab ends gracefully, robot.q reflects the final pose afterward