[CRYPTO] salsa20_generic: Fix multi-page processing

This patch fixes the multi-page processing bug that affects large test
vectors (the same bug that previously affected ctr.c).

There is an optimization for the case walk.nbytes == nbytes. Also we
now use crypto_xor() instead of adhoc XOR routines.

Signed-off-by: Tan Swee Heng <thesweeheng@gmail.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>