Since we had some problems with expr I rewrote the parsing logic, which then again ballooned into rewriting everything (as seems to happen to me often these days...).

I removed the tokenization, because that's not correct; each token could also be a valid string in another position. I also made the AST so that it does not have slices of values, but a fixed number for each operation. This makes the whole thing much less error prone. The parser is replaced with a fairly simple yet effective recursive descent parser.

The total implementation is also much smaller (especially considering that some lines are new tests).

Limitations:

• Currently I've only added String values, which leads to unnecessary conversions between strings and integers.
• Evaluation should happen during parsing not after (try for instance expr '(' 1 / 0, which in GNU gives a division by zero error and a syntax error here)