feat: spectator mode - all visitors see the same streaming generation

fix: move k_lens to GPU in SDPA fallback (tested locally)

fix: build proper attention mask in SDPA fallback for text cross-attention

perf: increase server frame buffer target to 16 for batch fetching

revert: remove float16 override, bfloat16 is safer (float16 overflow risk)

perf: use float16 instead of bfloat16 in SDPA for T4 tensor core acceleration

fix: adapt model_manager to HF model API (no schedule_config/cfg_config dicts)

fix: cast SDPA output back to original dtype + idempotent patching

fix: patch flash_attention with SDPA fallback for T4 (no flash-attn)

feat: initial FloodDiffusion streaming demo for HF Space