perf(lfm2-audio): extract shared backbone_step, TTS/S2S prefills use gallocr

Extract the ASR run_step lambda into a standalone lfm2_backbone_step()
function that returns both logits and hidden state. The TTS synthesize
and S2S prefills now call this function instead of duplicating the
260-line backbone graph with a 256 MB bump-allocated buffer.

This completes the gallocr migration for all user-facing paths:
- ASR decode: gallocr (already done)
- ASR prefill: gallocr via backbone_step
- TTS prefill: gallocr via backbone_step (was 256 MB bump)
- TTS decode: gallocr via step1 lambda
- S2S prefill: gallocr via backbone_step (was 256 MB bump)
- S2S decode: gallocr via step1 lambda

Remaining compute_meta users: 2 (run_lfm diff harness paths — test-only).

Eliminates ~500 lines of duplicated prefill code. All paths are now
GPU-compatible via gallocr + ggml_backend_graph_compute.

ASR output verified unchanged on JP Q4_K. 433 unit tests passing.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Author: crispasr integration