gpio: sysfs: use per-gpio locking

Add a per-gpio mutex to serialise attribute operations rather than use
one global mutex for all gpios and chips.

Having a single global lock for all gpios in a system adds unnecessary
latency to the sysfs interface, and especially when having gpio
controllers connected over slow buses.

Now that the global gpio-sysfs interrupt table is gone and with per-gpio
data in place, we can easily switch to using a more fine-grained locking
scheme.

Keep the global mutex to serialise the global (class) operations of gpio
export and unexport and chip removal.

Also document the locking assumptions made.

Note that this is also needed to fix a race between gpiod_export and
gpiod_unexport.

Signed-off-by: Johan Hovold <johan@kernel.org>
Reviewed-by: Alexandre Courbot <acourbot@nvidia.com>
Signed-off-by: Linus Walleij <linus.walleij@linaro.org>