Replace the ARCH_NR_GPIOS-sized static array of GPIO descriptors by
dynamically-allocated arrays for each GPIO chip.

This change makes gpio_to_desc() perform in O(n) (where n is the number
of GPIO chips registered) instead of O(1), however since n is rarely
bigger than 1 or 2 no noticeable performance issue is expected.
Besides this provides more incentive for GPIO consumers to move to the
gpiod interface. One could use a O(log(n)) structure to link the GPIO
chips together, but considering the low limit of n the hypothetical
performance benefit is probably not worth the added complexity.

This patch uses kcalloc() in gpiochip_add(), which removes the ability
to add a chip before kcalloc() can operate. I am not aware of such
cases, but if someone bisects up to this patch then I will be proven
wrong...
Signed-off-by: Alexandre Courbot <acourbot@nvidia.com>
Signed-off-by: Linus Walleij <linus.walleij@linaro.org>