Rework CAAM RNG implementation as follows:

- Make use of the fact that HWRNG supports partial reads and will
handle such cases gracefully by removing recursion in caam_read()

- Convert blocking caam_read() codepath to do a single blocking job
read directly into requested buffer, bypassing any intermediary
buffers

- Convert async caam_read() codepath into a simple single
reader/single writer FIFO use-case, thus simplifying concurrency
handling and delegating buffer read/write position management to KFIFO
subsystem.

- Leverage the same low level RNG data extraction code for both async
and blocking caam_read() scenarios, get rid of the shared job
descriptor and make non-shared one as a simple as possible (just
HEADER + ALGORITHM OPERATION + FIFO STORE)

- Split private context from DMA related memory, so that the former
could be allocated without GFP_DMA.

NOTE: On its face value this commit decreased throughput numbers
reported by

  dd if=/dev/hwrng of=/dev/null bs=1 count=100K [iflag=nonblock]

by about 15%, however commits that enable prediction resistance and
limit JR total size impact the performance so much and move the
bottleneck such as to make this regression irrelevant.

NOTE: On the bright side, this commit reduces RNG in kernel DMA buffer
memory usage from 2 x RN_BUF_SIZE (~256K) to 32K.
Signed-off-by: Andrey Smirnov <andrew.smirnov@gmail.com>
Reviewed-by: Horia Geantă <horia.geanta@nxp.com>
Cc: Chris Healy <cphealy@gmail.com>
Cc: Lucas Stach <l.stach@pengutronix.de>
Cc: Horia Geantă <horia.geanta@nxp.com>
Cc: Herbert Xu <herbert@gondor.apana.org.au>
Cc: Iuliana Prodan <iuliana.prodan@nxp.com>
Cc: linux-crypto@vger.kernel.org
Cc: linux-kernel@vger.kernel.org
Cc: linux-imx@nxp.com
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>