xen: pvblock: fix the maximum io size in one operation

The current implementation may cause BUG_ON() in blkfront_aio()
	BUG_ON(n > BLKIF_MAX_SEGMENTS_PER_REQUEST);

In pvblock_iop(), a read/write operation will be split into smaller
chunks of data so that the size in one access (aio_nbytes) is limited
to, at the maximum,
	BLKIF_MAX_SEGMENTS_PER_REQUEST * PAGE_SIZE

But this works only if when the *buffer* passed in to pvblock_io()
is page-aligned. If not, the given data region may stand across
(BLKIF_MAX_SEGMENTS_PER_REQUEST + 1) pages. See the logic in
blkfront_aio():
	start = (uintptr_t)aiocbp->aio_buf & PAGE_MASK;
	end = ((uintptr_t)aiocbp->aio_buf + aiocbp->aio_nbytes +
	       PAGE_SIZE - 1) & PAGE_MASK;
Then this will lead to BUG_ON() above.

This can be fixed by decreasing the maximum size of aio_nbytes.

Signed-off-by: AKASHI Takahiro <takahiro.akashi@linaro.org>
Fixes: commit 3a739cc6 ("xen: pvblock: Implement front-back protocol and do IO")