Skip wake word scoring while a pipeline/response is active

[jbeta4711]

Summary

• Satellite.wakeup() already ignores wake word detections while self._pipeline_active is True (logged as "Ignoring wake word - pipeline already active"), but the wake word models keep running inference on every audio frame during that window regardless, including throughout TTS playback.
• On hardware without working AEC, a satellite's own TTS output reliably scores as a near-certain match for its own wake word (residual echo picked up by the mic) - in real-world testing I saw probabilities of 0.9+ throughout entire responses, well above the configured threshold.
• wakeup() correctly catches this today, but only after the fact. There's also a narrow window right after REFRACTORY_SECONDS elapses mid-response (before _pipeline_active clears) where a fresh detection depends entirely on that one guard.
• This moves the check to the source: skip wake word scoring for the primary wake words while a pipeline/response is already active, matching the intent already documented in satellite.py:

  "When TTS is playing, keep _pipeline_active = True to block false wake word detections from speaker audio feedback."

• Stop word detection is a separate code path and is unaffected - it still needs to run while a response is in progress.

Test plan

• ./script/lint_black, ./script/lint_isort, ./script/lint_flake8 clean on the changed file (couldn't run the full ./script/tests suite locally - missing libmpv-dev system dependency in my sandbox, not something I could install there)
• Deployed to 5 real satellites (Raspberry Pi, ReSpeaker 2-Mic HAT) running microWakeWord, confirmed no probability-log lines for the primary wake words during active playback afterwards, containers stayed healthy
• Would appreciate a maintainer/CI run of the full test suite given I couldn't run it myself

🤖 Filed with the help of Claude Code, reviewed and approved by me before submission.