Optimize Chatterbox KV cache reads

[jltjarvinen]

Summary

Pass GPT-2 K/V cache views directly to ggml_flash_attn_ext instead of materializing the full KV history with ggml_cont() on every layer and decode step.

The naive attention fallback still materializes contiguous K/V tensors as before.

The same optimization can also be applied to the shared attention code in attention.h. The flash-attention path can consume the strided KV-cache views directly.

Result

On Chatterbox Nano CPU T3 decoding, this improves decode performance by about 2%

[CrispStrobe]

Merged after audit, with a follow-up proof on main.

The direct per-layer K/V views are valid: their logical layout is contiguous and only the layer view offset is nonzero; ggml_cont() therefore added a needless copy before flash attention. Follow-up f40e4427 adds a real CPU ggml graph regression comparing a nonzero-offset cache layer directly against the old materialized path (71 assertions), plus CRISPASR_CHATTERBOX_KV_CONT=1 for model-backed A/B.

Hosted live run 33333867440 is green: public Nano Q4_K + Turbo S3Gen Q4_K, same 40-token trajectory and byte-identical decoded PCM between direct-view and old-materialized paths. (The WAV containers themselves carry timestamped provenance, so the workflow correctly compares stream parameters and PCM rather than container bytes.)