Avoid some widening operations

[cuviper]

• The fallback implementations of adc and sbb can avoid widening by mimicking the standard library's implementations of carrying_add and borrowing_sub. (But the intrinsic x86/x86_64 versions are still faster.)
• The inverse used in Montgomery multiplication doesn't need to be wide, nor signed, so we can drop that and remove SignedDoubleBigDigit altogether.