Store quantized input/weight and their inverse scales in _Float8Matmul ctx to avoid re-quantization in backward and reduce saved-activation memory without changing numerics.
