Rewrite bitstream reader, enable partial decoding

[tirr-c]

This PR removes previous Read-based bitstream reader and instead adds memory-backed (&[u8]) reader. As a result the overall interface was rewritten around push-based approach, which users "push" (or "feed") bytes to the decoder, and then get rendered images. This API change has some advantages:

• The new API is async friendly. Users can asynchronously wait for some data, push them into decoder, then repeat. While waiting for data, they could progressively render images with the data given so far. (Decoding and rendering themselves are not IO-bound, so there's no need to change those interfaces to async.) This use case made me implement...
• ...improved partial decoding. Now jxl-render allows partial data for even more cases than before, like partial LF/HF coefficients, partial modular subimage of LfGroup, and even partial GlobalModular. Below is an example of this, loading ~379 KiB of JPEG XL image (conformance image bike) in chunks of 4 KiB:

bike.mp4