Keep MLX dense event-rate state on device

[cboulay]

Summary

• keep dense COUNT + SUM carry state in the input array namespace to avoid per-chunk MLX device synchronization
• reset threshold state after an empty first chunk so the next MLX chunk has a valid previous-sample reference
• preserve sparse threshold output by default and require explicit DENSE configuration for accelerator pipelines
• add NumPy and MLX regression coverage for output contracts, empty-first-chunk handling, and end-to-end dense event-rate processing

Testing

• uv run pytest -q (155 passed)
• uv run ruff check src tests
• git diff --check dev --