vLLM issue #50881: MiniMax-M3 fp8 KV cache on SM80 is coherent with --enforce-eager and garbage under CUDA graphs

Reporter supplied a working fix (@eager_break_during_capture on MiniMaxM3SparseAttention._insert_kv plus routing dense KV writes through reshape_and_cache_flash). Also reports SM80 e4m3 KV is refused outright ('type fp8e4nv not supported in this architecture') so only e5m2 is loadable there.