Fix various assumptions about mpw sizes.

  * configure, mptypes: New configure switches force mpw type to either
    sane but small (16/32 bits) or cussid (19/38 bits).  This found a
    bunch of exciting bugs...

  * gfreduce, mpreduce: If MPW_BITS is not a power of two, modular
    reduction of a `negative' unsigned value does the wrong thing.

  * mpx_lsl and friends: Shifting ops weren't masking high-order bits
    correctly when writing the output.  Apply MPW().

  * mpx_usubnlsl: More failure to elide high-order junk bits.

  * mptypes, mpx_udiv, mpx_bits, mp_odd: The binary search is neato, but
    starts in the wrong place if MPW_BITS is not a power of two.  Have
    mptypes  compute MPW_P2 as the largest power of two less than
    MPW_BITS.