I get the impression that you think of SPI as a faster version of standard (UART) serial interface. It isn't.
if you call
byte rv = SPI.transfer(cmd);
the rv byte is not the response of the slave to the cmd byte sent but it's the value that is received at the same time as the cmd byte is sent. So bit 7 of cmd is sent while bit 7 of rv is received and so on. That's why you need the delayMicroseconds() call (although 5µs would be probably enough to react on the received value and fill the data register).