docs(asr): document VRAM-vs-audio-duration limits and add chunked inf…

[AviArora02-commits]

Closes #367

Problem

docs/vibevoice-asr.md states "60-minute single-pass processing" without VRAM
qualification. On RTX 4090 (24 GB), default sdpa OOMs beyond ~30 min
(empirically: 30 min → ✅ ~22 GB peak; 50 min → ❌ OOM).

Changes

• docs/vibevoice-asr.md: Add Hardware Requirements section documenting
  the VRAM-vs-duration relationship and recommending flash_attention_2 for
  ≤24 GB GPUs.
• demo/vibevoice_asr_chunked_inference.py: Minimal chunked inference script
  for GPUs where flash-attn is unavailable.

Notes

• No model code modified.
• Chunked inference carries a known caveat (per-chunk diarization IDs are not
  globally consistent); this is documented inline.