Release 0.11.1: benchmark throughput and pipeline improvements

[dylanuys]

Bumps the package to 0.11.1 with faster video and audio preparation and shared benchmark pipelines.

• Sequential PyAV video decoding: reads the requested frames directly, avoiding whole-file copies and indexing. A matched, warmed 52-input workload improved from 33 → 23 seconds (~1.4×), with identical predictions.
• Parallel audio preparation: reusable decoder workers reduced a warmed workload from 139 → 93 seconds (~1.5×), with identical predictions.
• Lower checkpoint overhead: grouped writes improved a matched video workload from 262 → 115 seconds (~2.3×); streamlined persistence and recovery preserve integrity checks and explicit durable saves.
• Efficient input handling: metadata-based startup validation, augmentation cache reuse, and skipping checkpointed inputs avoid redundant work.
• Shared pipelines: consolidated runners, prefetching and cache handling; focused transform modules; stage timings for diagnosing remaining bottlenecks.

These are separate, small workloads; gains vary and are not additive. Video decoding accepts readable prefixes without requiring complete frame-count metadata or a readable tail. Existing checkpoints require a matching evaluator to resume.

Validation: automated tests, decoded-frame comparisons across 271 datasets, matched GPU predictions, interruption/recovery checks, and a clean package build.