feat: preserve raw input audio in frame userdata

[chenghao-mou]

Warning

This should not be merged until backend can handle the new channel

Expose aligned raw input through lk.audio.raw and label Krisp output with lk.audio.processing (denoised or isolated). Consumers select audio before resampling.

Pre-connect audio is resampled before processing so the raw audio and processing label remain available at agent input.

Tests cover raw/processed alignment with buffered and mutating processors, pass-through frames, and pre-connect resampling.

Addresses AGT-3508 (Linear).

Initial prompt and agent context

Model: GPT-6

for noise cancellation, can we preserve a raw field in userdata + processing=denoised vs isolated? if the plugin is NC, we keep raw + denoised, if it is VF, we keep raw + isolated $grill-me

immediately before. Resampling is fine

at input is enough, downstream resampling should be on the target audio frame. anyway.

FrameProcessor first

we should name it lk.audio.raw and lk.audio.processing

okay, let's create a PR to add this support, and also include a 3rd channel in recorder io to include the raw user input.

I guess we can skip recorder changes for now: Native Opus fallback loses recordings

When libopus is unavailable, CHANNEL_LAYOUT sends three channels to the native Opus encoder. That encoder accepts only mono or stereo, so encoding fails and leaves the recording incomplete.