Fix FixedDecimal rounding bug; add more tests

[sffc]

Numbers with interior zeros were expanding at too high a magnitude. For example, consider the input 1.108 expanded at position -2.

Old behavior:

1. Remove the digits beyond position -2 to get 1.10 (digits list 110)
2. Normalize the internal representation (digits list 11)
3. Increment the last digit to get 1.20

New behavior:

1. Remove the digits beyond position -2 to get 1.10 (digits list 110)
2. Increment the last digit to get 1.11
3. Assert that the internal representation is normalized

I had the individual failing input from @Manishearth, but I had to write some elaborate test cases before I could reproduce the problem from "first principles". Since I was able to do that, I'm confident that this change fixes the problem across all rounding modes.

[Manishearth]

cc @eggrobin who was also observing interestingly as I was poking at this failure

[sffc]

I went ahead and refactored the logic of the trunc and expand functions to be more clear and consistent and avoid calling into other helper functions that obscured the behavior.