[ARM] Fix some corner cases in new mm initialisation

Document that the VMALLOC_END address must be aligned to 2MB since
it must align with a PGD boundary.

Allocate the vectors page early so that the flush_cache_all() later
will cause any dirty cache lines in the direct mapping will be safely
written back.

Move the flush_cache_all() to the second local_flush_cache_tlb() and
remove the now redundant first local_flush_cache_tlb().

Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>