Even if well documented, this is rather unusual to have read function mutating the input to be read. Is there a real benefit in doing so vs returning a shorten copy of the input bytes ?

The benefit is that you can write really neat code parsing structured data:

fn parse_slpv2(data: &mut Bytes) -> Result<...> {
  let lokad_id: [u8; 4] = read_array(data)?;
  let token_type = read_array::<1>(data)?[0];
  let tx_type_len= read_array::<1>(data)?[0];
  let tx_type = read_bytes(data, tx_type_len as usize)?;
  // ...
}