Bug in NNPF floating-point inference

• Typo and casting missing on input/output values formatting
• Check of output tensor quantisers wrt out_tensor_luma_bitdepth_minus8 is done even when float inference is used