Refactor _decode_labels_at_offset to avoid recusion

[bdraco]

_decode_labels_at_offset is the most expensive function in terms of execution overhead in the whole stack

part of that is the python overhead of the recursion

[bdraco]

[bdraco]

RPI3

[bdraco]

Its the python call overhead here that is the issue

[bdraco]

refactoring didn't help much. This is just expensive unfortunately.

[bdraco]

I really don't see how this can be made faster unless we make a cython extension. We could do pxd files so it would work with or without, but then we need to publish wheels

[bdraco]

incoming parsing and _dns object creation even public would get a major boost if we did that

[bdraco]

I couldn't get rid of the recursion but I did get a 36% speed up with #1097

[bdraco]

The number of re-entries is significantly reduced after #1097

[bdraco]

Any future improvements would likely need to be a C extension, but 35% improvement helps a bit.

The RPI3b is still a bit unstable at startup but any little bit helps