dm: ns16550: Restore how we define UART_REG

dm: ns16550: Restore how we define UART_REG

Prior to commit 9591b63531fa ("Convert CONFIG_SYS_NS16550_MEM32 et al to
Kconfig") we had defined CONFIG_SYS_NS16550_REG_SIZE to -1 with
DM_SERIAL such that we would then have a size 0 character array. This
resulted in functionally no padding. The confusion on my part came from
dealing with the constraints around platforms that do not use DM_SERIAL
in SPL/TPL. After Andre Przywara reported that sunxi was broken, I've
re-read the code and comments again and thought on this harder. What we
want I believe is what this patch does now.

If DM_SERIAL is defined for this stage, regardless of
CONFIG_SYS_NS16550_REG_SIZE then we will dynamically handle reg shifts
and 'struct ns16550' needs no padding (which is functionally what
unsigned char foo[0] provides). This is the same case as NS16550_DYNAMIC
and DEBUG_UART. Expand the existing comment here slightly.

Otherwise, we will have CONFIG_SYS_NS16550_REG_SIZE set to a non-zero
value, and handle padding within the struct.

Cc: Simon Glass <sjg@chromium.org>
Cc: Sergei Antonov <saproj@gmail.com>
Cc: Quentin Schulz <quentin.schulz@theobroma-systems.com>
Cc: Andre Przywara <andre.przywara@arm.com>
Fixes: 9591b63531fa ("Convert CONFIG_SYS_NS16550_MEM32 et al to Kconfig")
Signed-off-by: Tom Rini <trini@konsulko.com>
Reviewed-by: Andre Przywara <andre.przywara@arm.com>
Tested-by: Andre Przywara <andre.przywara@arm.com>
Tested-by: Quentin Schulz <quentin.schulz@theobroma-systems.com>