ggml-cuda: skip duplicate decode sibling prefetch

[GenerelSchwerz]

Summary

Skip redundant sibling prefetch passes during single-token MoE decode after the routing-ID cache identifies later siblings. Preserve existing prefill behavior.

Validation

• test-moe-cache: passed
• Untouched dbeb350d1 and candidate PPL: 1.0018 +/- 0.00011
• Deterministic raw token IDs and token accounting matched base at 8K and 32K
• Cache misses, evictions, H2D copies, and VRAM matched base
• 48-slot decode: +0.24% at 8K, +0.49% at 32K
• Equal-VRAM decode: neutral at 8K/188 slots and 32K/184 slots
• Combined with routing-ID reuse at 0c5347236, prefill telemetry matched base and 48-slot decode was +0.31% at 8K and +0.15% at 32K
• All server lifecycles stopped cleanly

AI assistance: Codex assisted with implementation, profiling, and validation.