instead of left-to-right. This cuts the time for exponent 2
by another 10%. Does that slow things for, e.g., an exponent
like 1 << 20? Sure, by a little.  But that exponent will go
on to do 20 squarings, and the extra native 32-bit shift-
and-tests in the loop are insignificant in comparison. For
a tiny exponent (like 2), the loop's shift-and-tests can
consume a significant part of the total time.

Also reverted the change I made to repair refcounting in
modular inverse code. Mark will make the change in a
distinct PR, and backport it. Without this change, the
"negative refcount" glitch stopped showing up in this PR's
code after I changed it to never return `a` directly.

until the exponent is at least 4. Saves approaching another 10%
for exponent 2.

insteads works OK, but was always a conceptual mess.

Close to another 10% saved for exponent 2.