Title: Gate EMA accumulator on data_ena for multiplexed-stream use cases

[Abraxas3d]

Summary

Adds a data_ena = '1' gate around the EMA accumulator update inside lowpass_ema. data_ena now selects which clocks the EMA integrates, rather than only riding through as a passsthrough strobe to average_ena.

Motivation

lowpass_ema previously updated all internal state every clock and forwarded data_ena unchanged to average_ena. This is correct for streaming inputs (data valid every clock, data_ena tied high), but incorrect for any use case where data_ena selects which clocks have valid samples (like Haifuraiya).

We discovered this while building a 64-channel polyphase-channelizer power monitor for Haifuraiya. The wrapper instantiates 64 power_detector blocks sharing a single multiplexed I/Q data bus; each instance's data_ena selects which clocks correspond to its assigned channel. Without the data_ena gate, all 64 power_detector instances accumulated the same shared data stream and reported identical averages, regardless of which channel each one was supposed to monitor. Took a while to figure this out!

Change

Wraps the existing accumulator assignments in if data_ena = '1' then ... end if; inside the else (non-reset) branch of the clocked process. average_ena <= data_ena; remains outside the gate so output validity is reported faithfully regardless of accumulator update.

Backward compatibility

For streaming inputs with data_ena tied high, behavior is unchanged: the gate is always satisfied and every clock updates as before.

For users who tied data_ena permanently low, the EMA now holds state instead of overwriting with the input alpha/data. We believe this is the correct behavior (no data, no update) and are not aware of any user relying on the previous behavior. pluto_msk does not.

Testing

Verified in the Haifuraiya AXI wrapper testbench:

• DC input (20000 + 0j) -> peak channel 0, with symmetric skirt
  decaying through channels 63, 1, 2, 62, 3, 61, ... and deep
  stopband in channels 17-46. Pre-fix: all 64 channels read
  identical EMA values.
• Tone at FFT bin 16 -> peak channel 16, with symmetric skirt
  around bin 16. Pre-fix: all 64 channels read identical EMA
  values.

Addendum!

Independently caught during the same session: the README port table states power_squared width is 2*DATA_W - 2 but the actual range declaration (2*DATA_W - 2 downto 0) is 2*DATA_W - 1 bits wide.