Improve performance, especially in data with many CR-LF

[jhnstrk]

The code changes here started as an attempt to improve performance with data loaded with many CRLF pairs (#67). In the current code these cause performance issues because CRLF matches the start of a boundary pattern. This makes advancing through the data very slow due to the number of call-backs.

By far the greatest optmization was to remove the partial Boyer-Moore-Horspool (BMH) implementation and replace it with bytes.find and a bit of logic to ensure partial matches at the end weren't missed. bytes.find appears to be significantly faster (around 3 times) on representative data. I can make a pull request with just this change if you are interested.

In the update I also attempted to reduce the number of times the on_part_data callback was called to a minimum. Whereas the old code would call it every time a partial boundary match was found (i.e. CRLF), now it only calls it when necessary. The conditions for calling the on_part_data are now:

1. A complete boundary match is found, either for an end of part or final end.
2. The currently loaded data buffer has been exhausted.

The significant complication is what happens when a partial boundary match overlaps the end of the loaded data. This was addressed with a look-behind buffer before, but the buffer is mostly unnecessary: since we are always matching boundary bytes, the look-behind buffer is always just a copy of the boundary. Only the last few bytes may vary (CRLF vs -- depending on whether it is a part or end boundary). However hitting this condition should be very very rare, and is addressed in the code.

[Kludex]

I can make a pull request with just this change if you are interested.

All good.

Thanks for the great description, helped a lot on the review.