We are getting randconfig build errors on device drivers with
tristate Kconfig option if they are using custom initcall
levels. Rather than add ifdeffery into the drivers, let's add
the missing initcall_sync variants.

As the comment in init.h has kept people from updating the
list of initcalls that can be just module_init when the driver
is loaded as a loadable module, let's also update the comment
a bit to describe valid use cases custom initcall levels.

While most drivers should nowadays use just regular module_init
because of the deferred probe, we do have quite a few custom
initcall levels left that we cannot remove until tested properly.

There are also still few valid cases where a custom initcall
level might make sense that I'm aware of.

For example a bus snooping driver can provide information about
invalid bus access and is handy loader early when built in. But
there's no hard dependency to have it necessarily built in and
a loadable module is a valid option.

Another example is a driver implementing a Linux framework like
pinctrl framework. That driver may be needed early on some
platforms because of legacy reasons, while it can be just a
regular module_init on most platforms.

Signed-off-by: Tony Lindgren <tony@atomide.com>
Cc: Arnd Bergmann <arnd@arndb.de>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>