My feeling is that it doesnt improve the situation:

Fragmentation may looses ram:
http://yarchive.net/comp/linux/page_sizes.html

mb/s isnt latency:
If your software needs random read like swap, you cant convert it to sequential by just changing the kernel page size.
(write isnt that important cause you can change the swappiness to swapout ahead of time (at the cost of battery life), the kernel first overwrites code pages which he just rereads from the filesystem (no need to write them, they are already saved) and so most swap operations are random reads (which are slow even compared to sequential(!) writes which you can do when you actually do swap out other data) (But this may depend on your swap policy))

I hope swap gets already paged in and out in multiple Blocks, no need to recompile:
http://www.gluster.org/community/doc..._Kernel_Tuning
"cat /proc/sys/vm/page-cluster"

But maemo + kernel seems so old (and sadly not good at io) that i found sparse, outdated and even contradicting info.

ram via sd would be nice
But not as nice as better software.

But i would like to be proofed wrong. So post real world experience!