Make the I/O faster, mainly by using combined SMBus transactions when
possible. While the TSL2550 datasheet doesn't say the device supports
them, they seem to work just fine in practice, and a combined
transaction is faster than two simple transactions in many cases and
always more reliable.

A side effect is to suppress the delays between SMBus writes and
reads. The datasheet doesn't say they are needed and things work just
fine for me without them.

I also couldn't see any reason for the delay between reading the two
channels. Nor for the loop to get a reading in the first place. The
400 ms delay between samples only matters at chip power-up, after that
the chip always hold the previously sampled value so we never get to
wait.

All these changes make reading the lux value much faster and cheaper.
Signed-off-by: Jean Delvare <khali@linux-fr.org>
Tested-by: Michele De Candia <michele.decandia@valueteam.com>
Cc: Rodolfo Giometti <giometti@linux.it>