Rewriter can now inform OutputSink what encoding is being used. Without that, the OutputSink had no way of correctly handling varying character encodings other than reparsing HTML itself.

Directly using interior-mutable SharedEncoding made it harder to reason about currently used encoding. With encoding change being an explicit step, a simpler AsciiCompatibleEncoding can be used in more places.

I've also replaced ALL_ENCODINGS lookup table with OnceLock. The encoding can be changed only once, so we can store AsciiCompatibleEncoding directly without going through AtomicUsize.