Optimize and simplify BitReaderReversed

[fintelia]

This PR refactors the BitReaderReversed object based on the structure from the Reading bits in far too many ways blog series.

It retains the get_bits method to read bits from the stream, but also exposes lower-level peek_bits, consume, and refill methods. These new methods may unlock optimizations in other parts of the code by reducing the number of hard-to-predict branches on whether to refill the bit container. And a separate peek_bits method might be especially useful for Huffman decoding because it lets you do a table lookup with N bits and then decide to consume fewer than N of them, rather than the current strategy of intentionally reading past the end of the input.

On my machine (a Ryzen 5600X), I see a ~5% end-to-end performance improvement in decode time for enwik9 from this change.

[KillingSpark]

This is an interesting changeset. I can't reproduce the performance gains on a Ryzen 5 3500U (it performs basically the same) but think I see the potential for the optimizations you mention in other regions of the library.

I'll look into this more after the holidays :)

Edit: Just for the simplification this is probably worth pulling, even if it doesn't improve performance just yet.

[KillingSpark]

Yep looks good, thanks for the PR!

Edit: On my Ryzen 3800X that's approximately a 10% improvement!