Streaming performance: kotlinx-io & Okio integrations

[qwwdfsad]

See #2742 for the benchmarks.

Currently, decoding from the buffered source using kotlinx-serialization is unreasonably slow compared to Moshi.

One of the reasons is that we do not leverage select API that potentially requires #2240 to be implemented first.

But even with this API out of the equation, we still have a single bottle-neck -- byte-by-byte codepoint reading (see the profile attached), and that definitely can be more straightforward and efficient (e.g. using raw access to the underlying segments).