1. Add a dump() method for faster debugging.

2. Change 0 initialization of union to larger component so all is zeroed.
3. Fix the borrow logic in subtraction so it works for > 128 bits.
4. Rewrite fromString to use a simpler but correct algorithm and also to
   not set the bit width contrary to the user's request.
5. Optimize toString a bit by making it only do one Knuth divide per
   iteration instead of two.

With these changes, all arithmetic passes (verified by pari/GP) up to
1024 bits except for certain division cases.

llvm-svn: 34463