atmel_spi: chain DMA transfers

Add support for chained transfers in the atmel_spi driver, letting the DMA
controller switch to the next buffer pair without CPU intervention.  This
reduced I/O latencies by about 2% in one bulk I/O test.  It should also help
work around several interrelated errata affecting chipselect 0 on at91rm9200
chips.

Almost all of the changes are in the reworked atmel_spi_next_xfer() function.
That's now called with the driver in one of three states:

 1. It isn't transferring anything (in which case the first transfer
    of the current message is going to be sent)
 2. It has finished transfering a non-chainable transfer (in which
    case it will go to the next transfer in the message)
 3. It has finished transfering a chained transfer (in which case the
    next transfer is already queued)

After that it will queue the next transfer if it can be chained.

Signed-off-by: Szilveszter Ordog <slipszi@gmail.com>
Acked-by: Haavard Skinnemoen <hskinnemoen@atmel.com>
Signed-off-by: David Brownell <dbrownell@users.sourceforge.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>