Add marvell_nand driver which aims at replacing the existing pxa3xx_nand
driver.

The new driver intends to be easier to understand and follows the brand
new NAND framework rules by implementing hooks for every pattern the
controller might support and referencing them inside a parser object
that will be given to the core at each ->exec_op() call.

Raw accessors are implemented, useful to test/debug memory/filesystem
corruptions. Userspace binaries contained in the mtd-utils package may
now be used and their output trusted.

Most of the DT nodes using the old driver kept non-optimal timings from
the bootloader (even if there was some mechanisms to derive them if the
chip was ONFI compliant). The new default is to implement
->setup_data_interface() and follow the core's decision regarding the
chip.

Thanks to the improved timings, implementation of ONFI mode 5 support
(with EDO managed by adding a delay on data sampling), merging the
commands together and optimizing writes in the command registers, the
new driver may achieve faster throughputs in both directions.
Measurements show an improvement of about +23% read throughput and +24%
write throughput. These measurements have been done with an
Armada-385-DB-AP (4kiB NAND pages forced in 4-bit strength BCH ECC
correction) using the userspace tool 'flash_speed' from the MTD test
suite.

Besides these important topics, the new driver addresses several
unsolved known issues in the old driver which:
- did not work with ECC soft neither with ECC none ;
- relied on naked read/write (which is unchanged) while the NFCv1
  embedded in the pxa3xx platforms do not implement it, so several
  NAND commands did not actually ever work without any notice (like
  reading the ONFI PARAM_PAGE or SET/GET_FEATURES) ;
- wrote the OOB data correctly, but was not able to read it correctly
  past the first OOB data chunk ;
- did not retrieve ECC bytes ;
- used device tree bindings that did not allow more than one NAND chip,
  and did not allow to choose the correct chip select if not
  incrementing from 0. Plus, the Ready/Busy line used had to be 0.

Old device tree bindings are still supported but deprecated. A more
hierarchical view has to be used to keep the controller and the NAND
chip structures clearly separated both inside the device tree and also
in the driver code.

Signed-off-by: Miquel Raynal <miquel.raynal@free-electrons.com>
Tested-by: Sean Nyekjaer <sean.nyekjaer@prevas.dk>
Tested-by: Willy Tarreau <w@1wt.eu>
Signed-off-by: Boris Brezillon <boris.brezillon@free-electrons.com>