Audio: add payload-neutral global bucketing

[mohammadaaftabv]

Summary

• plan model-safe audio segments and duration buckets from manifest metadata without loading waveforms
• preserve planner-owned batches as atomic dispatch envelopes through FastConformer inference
• decode bounded segments lazily for standalone operation and accept preloaded waveforms through a payload-neutral seam
• restore parent rows through a single-copy metadata repository with retry-safe distributed assembly

Independence

This PR is based on main and has no PayloadRef, payload-lifecycle, Qwen-Omni, or performance dependency. It runs standalone and can merge before or after the payload-lifecycle PR.

Test plan

• Ruff check and format validation
• Batch policy, segmentation, dispatch, decode/resampling, and assembly tests
• FastConformer dispatch-owner proof
• Distributed parent repository and retry/idempotency coverage
• Both merge orders produce identical combined trees
• Payload-hydrated end-to-end composition proof
• Live FastConformer GPU checkpoint run

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.