fix(qwen): preserve compact logits indices across graph reuse

[mirek190]

Summary

Fix compact-logit row-index tensors in the shared Qwen causal decode runtime being overwritten when a cached prefill graph is reused.

The compact readback path creates an I32 tensor consumed by ggml_get_rows. It is populated by the host after graph allocation, but it was not marked as a graph input. The allocator could therefore alias its storage with an intermediate tensor. A first request could succeed while leaving corrupted indices for the next execution, which was observed as an illegal CUDA memory access on the second request.

Changes

• mark compact-logit token IDs as a persistent ggml input
• re-upload the IDs immediately before every cached prefill execution
• cover both single and batched prefill paths
• retain cached graph reuse rather than rebuilding compact-logit graphs

Validation

• Windows CPU full-composite audiocpp_cli build: passed
• Windows CPU engine_core rebuild: passed
• Windows CUDA 12.4 engine_core build for SM 86: passed
• git diff --check: passed

This is isolated from the MiraTTS model work and changes only the shared Qwen runtime lifecycle for configurations using logits_readback_token_ids.