Use setjmp + longjmp to handle errors

[k0ekk0ek]

There are not that many places errors can occur and the library itself is designed to not allocate additional memory (except for file buffers). Handling errors in one place reduces the number of possible branches and therefore may positively impact performance and code size.

[k0ekk0ek]

I have an implementation of this locally. There's a slight increase in performance and code complexity is reduced too. Likely ends up in the work done for #19.