use the type's real alignment for the memmove/memcpy intrinsics, not 1

[thestinger]

LLVM carries this all the way down to the IR after optimization, so it's not using optimal instructions for swapping when it can't infer an alignment after inlining.

[thestinger]

This is fixed.