Patch looks correct and looks good to me, modulo a couple of nitpicks (see Rietveld comments).  This seems like a nice cleanup.

The patch introduces a new dependence on PY_UINT32_T, which is something we haven't so far used elsewhere in Python beyond the float<->string conversions.  That's quite a big change: it means that it's conceivable that the random module could now fail to build on platforms where it used to work.

It also changes the signature of 'init_by_array', which is one of the core routines taken directly from the MT implementation.