fix(mod_conference): prime conference input buffers

[Aglcr7]

Closes #3024.

Summary

• replace the mixer-side last-frame replay/concealment workaround with input-buffer priming
• keep a small real-audio cushion in each member audio_buffer before the conference mixer starts consuming it
• when a primed input buffer briefly drains below one complete frame, wait 1 ms and retry a real read before marking it unprimed
• mark the buffer unprimed again if it still cannot provide a complete real frame, so it re-primes from the input thread instead of replaying synthetic audio
• clear and unprime the input buffer when member media is configured/reset

Why

The issue is an input-thread underflow: the mixer can become phase-locked to a member input thread and sample the buffer just before the next real frame arrives. The previous approach concealed that by replaying the last frame, but that can corrupt FSK tones because it synthesizes audio. This version avoids replay and only mixes frames that came from the input thread.

Testing

• git diff --check
• verified no remaining audio_underflow / replay helper references in mod_conference

Full compile/test was not available in my Windows workspace because gcc, clang, cl, make, and cmake were not present in PATH.

[figaro2015]

We had a similar fix implemented already. Unfortunately, this fix won't work for the FSK tones. The real fix should be eliminating the actual input thread underflow condition.

[Aglcr7]

Thanks @figaro2015, I reworked the patch around that point.

The updated commits (4eed0b2, refined by 1e1fb29) remove the last-frame replay/concealment path and instead prime each member input audio_buffer with a small real-audio cushion before the mixer consumes it. If a primed buffer briefly drains below a complete frame, the mixer gives the input thread a 1 ms grace period and retries a real read; if no complete real frame arrives, the buffer is marked unprimed and has to re-prime from the input thread again. That keeps the fix on real input frames and avoids synthesizing/repeating samples, which should be compatible with FSK tones.

Local checks: git diff --check, plus verified the old audio_underflow / replay helper references are gone. I still cannot compile in this Windows workspace because the C toolchain is unavailable here.