Rewrite decode_performance_opt.py from empty stats shell to real optimization:
1. FlashAttentionBackend.forward fast-path (x64/step):
   - Cache kv_cache.unbind(0) per step (shared across 64 layers)
   - Skip redundant assert/attn_type/fp8/dcp checks for decoder-only models
   - Cache descale_shape and sliding_window list
   - Safe fallback: validates assumptions on first call, falls back permanently
     if not decoder-only/non-fp8/single-GPU; try/except on every call
2. _get_cumsum_and_arange uniform-decode fast-path:
   - Skip cumsum+repeat when all requests schedule 1 token
3. Lightweight decode/prefill step counter

Compliance: no scheduler logic change, no locked params, no semantic change.
Optimized forward produces identical output as original (only caches invariants).