mtd: mtdpart: Make mtdpart's _erase method sane

mtd: mtdpart: Make mtdpart's _erase method sane

The _erase() method of the mtdpart driver, part_erase(), currently
implements offset shifting (for given mtdpart partition) in a weird way:
  1. part_erase() adds partition offset to block address
  2. parent driver's _erase() method is called
  3. parent driver's _erase() method calls mtd_erase_callback()
  4. mtd_erase_callback() subtracts partition offset from block address
     so that the callback function is given correct address
The problem here is that if the parent's driver does not call
mtd_erase_callback() in some scenario (this was recently a case for
spi_nor_erase(), which did not call mtd_erase_callback() at all), the
offset is not shifted back.

Moreover the code would be more readable if part_erase() not only added
partition offset before calling parent's _erase(), but also subtracted
it back afterwards. Currently the mtd_erase_callback() is expected to do
this subtracting since it does have to do it anyway.

Add the more steps to this procedure:
  5. mtd_erase_callback() adds partition offset to block address so that
     it returns the the erase_info structure members as it received them
  6. part_erase() subtracts partition offset from block address

This makes the code more logical and also prevents errors in case
parent's driver does not call mtd_erase_callback() for some reason.

(BTW, the purpose of mtd_erase_callback() in Linux is to inform the
 caller that it is done, since in Linux erasing is done asynchronously.
 We are abusing the purpose of mtd_erase_callback() in U-Boot for
 completely different purpose. The callback function itself has empty
 implementation in all cases in U-Boot.)

Signed-off-by: Marek Behún <marek.behun@nic.cz>
Reviewed-by: Simon Glass <sjg@chromium.org>
Tested-by: Masami Hiramatsu <masami.hiramatsu@linaro.org>