Improve HackRF streaming throughput and recover unvalidated SDR tuning

[plaingca]

HackRF users reported TX underruns and repeated RX queue overruns across V7, V8 and AC16. Lower One's sample clock from 9.6 to 8.064 MS/s, optimize host filtering/oscillator/IQ packing, and prime the bounded TX queue before USB startup. One lacks Pro-style filtered FPGA decimation; the new clock remains above the recommended 8 MHz minimum.

The RC tester now reports no queue overruns or host TX underruns, but an unrelated HT transmission could leave RX searching until restart. Expire an unvalidated coarse frequency correction after 30 seconds; validated modem payload refreshes the deadline before neural decoding. Reset downstream acquisition and A/V state on retry. Record optional firmware shortfall counters at TX close to distinguish device-side issues from host queue underruns.

Validation: recovery full suite 492 passed, 2 skipped; after optional MCU diagnostics, focused HackRF/SDR suite 45 passed. Local DSP-only benchmarks show 70% lower TX cost and 62% lower RX cost. Paired five-mode quality simulation decoded every GOP with PSNR changes below 0.006 dB. All 13 user TX recordings replayed successfully (244/244 GOPs), including all five prepared clips. The recorded unrelated RX audio yielded no false payloads and then recovered all 29 GOPs when followed by valid recorded TX audio.

Prepared-clip stuttering heard on the tester's HT remains unresolved beyond the waveform-recording point: recorded waveform completeness does not establish USB/RF continuity. No local physical HackRF was available. See docs/hackrf-throughput.md and aggregate evidence; raw user recordings remain local. Evaluation build: v0.1.21rc2.