Add CUDA warmup inference on startup

[wpflueger]

Summary

• Runs a 1-second silent audio transcription after model loading to pre-compile CUDA kernels
• Reduces first-transcription latency (CUDA JIT compilation, memory allocation patterns)
• Adds ~1-2s to startup time; failure is non-fatal (logged as warning)
• Uses the model's own cfg.sample_rate to generate the correct length of silence

Test plan

• Verify "Running CUDA warmup inference..." and "CUDA warmup complete" appear in backend logs during startup
• Compare first-dictation latency with and without warmup
• Verify startup still succeeds if warmup fails for any reason

Closes #31

🤖 Generated with Claude Code

[Copilot]

Pull request overview

Adds a startup warmup inference step after loading the NVIDIA NeMo Parakeet model to reduce first-request latency caused by CUDA JIT/kernel initialization.

Changes:

• Calls a new _warmup() method immediately after self._model.eval() in ModelLoader.load().
• Implements _warmup() by running a 1-second silent transcription using self._model.cfg.sample_rate, logging success and treating failure as non-fatal.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

Warmup is invoked unconditionally after loading, but the loader allows device to be configured (e.g., PARAKEY_DEVICE=cpu). In that case this adds startup latency and logs misleading "CUDA warmup" messages without actually warming CUDA. Consider gating warmup on self._device == "cuda" (and CUDA availability) or renaming the log messaging to a device-agnostic warmup when not on CUDA.

[Copilot]

The warmup path calls self._model.transcribe(audio=[warmup_tensor]) with only the torch-tensor variant. In transcribe() you already implement compatibility fallbacks for NeMo versions that don’t accept tensors (or require positional args). If warmup hits one of those cases it will always fail and the intended CUDA kernel warmup won’t happen. Consider reusing the same fallback strategy here so warmup remains effective across supported NeMo/PyTorch combinations.

Suggested change

	self._model.transcribe(audio=[warmup_tensor])
	try:
	# Preferred path: newer NeMo versions accepting torch tensors
	self._model.transcribe(audio=[warmup_tensor])
	except TypeError:
	# Fallbacks for NeMo versions with different transcribe signatures
	try:
	# Some versions expect positional arguments
	self._model.transcribe([warmup_tensor])
	except TypeError:
	# Some versions may not accept tensors at all; try NumPy input
	self._model.transcribe([warmup_audio])