Fixes #440

The current logic for detecting overflows while parsing large integer only detects some overflows. Each calculation like multiplication and addition has to be verified to be in bounds because otherwise it might silently overflow.

This PR fixes overflow detection and out of bounds checks for integer numbers with large absolute value. It replaces the current bounds checking and overflow checking logic.
The implementation should be quite efficient because the checks only use additions and comparisons of 64bit integers in the parsing loop and it handles negative numbers without additional checks. Furthermore it correctly parses LLONG_MIN on 2s complement architectures, where -LLONG_MIN - 1 = LLONG_MAX should be satisfied. I haven't added a test for this because there is no guarantee that it will work on all architectures, i.e. -LLONG_MIN = LLONG_MAX is also possible. Maybe a test should be added because basically all CPUs use 2s complement representation of signed integers.

PR #543 seems to use a similar idea, but I only noticed it after writing the patch. It uses divisions in the parsing loop and I'm not sure whether it would parse all negative integers correctly.

Changes proposed in this pull request:

• add tests which fail with the current implementation
• fix overflow detection while parsing integer numbers
• make error messages consistent when parsing too large or too small numbers
• change the formatting of head of function