Rework magic data once more

The change reduces the input size on patterns using posix character
classes. It is still better than reverting to the patterns from the
older version, as the upstream data uses posix classes also in the later
versions. The input reduction speeds up the pattern matching in some
cases.

This patch is functionally almost same as upstream, but might show some diff
when the input is very long. While the magic data in the newer versions is
functionally an improvement, with jit=0 it might show a performance regression.
The slowdown is negligible in the normal usage and is still acceptable for the
malicious input. If some functional regressions show up, this patch should be
reverted and the tests timing should be adapted instead.

Author: weltling