[Bug] Fix fp8 bugs: 1. support only fp8 quant gemm or grouped gemm 2. precompute float8 dynamic scale before compute_actor_logprobs during rl training#1356