feat: SLAM streaming, recording, and playback (issues #78, #82)

[MoveFastAndBreakThings-dot]

What this does

Implements the full SLAM data pipeline between Shepard and Emu.

Changes

Streaming (issue #82)

• SLAMProvider — abstract base class defining get_frame() → SLAMFrame
• SLAMFrame — holds 3D point cloud (N×3 numpy), drone pose (x/y/z/roll/pitch/yaw), and timestamp
• SLAMEmuStreamer — background thread that pushes slam_data frames over WebSocket continuously
• OakdAltimeterSLAMProvider — real provider: fuses OAK-D depth camera + XM125 radar altimeter + MAVLink orientation
  into a georeferenced point cloud
• DebugSLAMProvider — fake point cloud for testing without hardware
• Emu.register_slam_streamer() — listens for {"type":"slam","command":"start/stop"} from Emu and starts/stops the
  stream to save bandwidth

Recording & Playback

• SLAMRecorder — transparent decorator around any SLAMProvider; records every frame to disk as compressed numpy
  (.npz) while passing frames through unchanged
• SLAMPlaybackProvider — replays a recording at original speed using preserved timestamps; implements SLAMProvider
  so it plugs directly into SLAMEmuStreamer
• RecordingManager — manages multiple named recordings (list, load, delete)

Samples

• slam_stream.py — test streaming with fake data (no hardware)
• slam_stream_real.py — real hardware (Pi + OAK-D + XM125)
• slam_record.py — record a session, Ctrl+C to stop and save
• slam_playback.py — replay any saved recording through Emu

Testing

All acceptance criteria verified by 8/8 automated tests:
PYTHONPATH=".:dep/labeller" python tmp/test_slam_local.py

End-to-end tested manually: slam_record.py streams to Emu SLAM tab live, recording saved to disk, slam_playback.py
replays identical session.

Notes

• recordings/ and tmp/ added to .gitignore — recordings stay local, not committed
• Real SLAM algorithm (issue Building the SLAM Pipeline #81) plugs in by implementing SLAMProvider.get_frame()