vLLM issue #47847: disable_flashinfer_q_quantization + fp8 KV + spec decode silently corrupts sliding-window models on SM100

Reporter root-caused the kernel defect in flashinfer-ai/flashinfer#3864 (trtllm_batch_decode_with_kv_cache with BF16-Q + FP8-KV mis-computes sliding-window attention past the window at certain seq_len/page alignments) and separately identified a q_scale double-application in vllm/v1/attention/backends/flashinfer.py that is inert when all scales are 1.0.