Release 0.11.0: audio robustness and benchmark improvements

[dylanuys]

• Release 0.11.0 with opt-in audio robustness: −6 dB gain, 30 dB SNR white noise, and an 8 kHz downsample/upsample round trip, using shared scoring, seeded caching, and resume support.
• Skip video decoding on robustness cache hits while preserving input validation.
• Preserve each dataset’s full-run sample limit when filtering datasets.

Validation: CI.