The original code was based on my reverse engineering of an I2C trace
of the Windows driver. Now that I know what the registers actually do,
restructure the code a bit, removing some unneeded register programming
and fixing the sequencing of operations.
This reduces the time it takes to change inputs from 1300ms down to
600ms (as measured by "time v4l2-ctl -i 0")
Note this does not address outstanding issues related to the management
of the module clocks and power control for the various blocks, which
will be done in a separate patch.