When setting up to transmit, a race exists between the ISR and
i2c_davinci_xfer_msg() trying to load the first byte and adjust counters.
This is mostly visible for transmits > 1 byte long.

The hardware starts sending immediately that MDR is loaded. IMR trickery
doesn't work because if we start sending, finish the first byte and an
XRDY event occurs before we load IMR to unmask it, we never get an
interrupt, and we timeout.

Move the MDR load after DXR,IMR loads to avoid this race without locking.

Tested on DM355 connected to Techwell TW2836 and Wolfson WM8985
Signed-off-by: Jon Povey <jon.povey@racelogic.co.uk>
Signed-off-by: Ben Dooks <ben-linux@fluff.org>