Streaming: startChunkedConversion / Stream.transform

[dipu-bd]

Today there is no chunked conversion — the base Converter inherits the SDK's startChunkedConversion, which throws UnsupportedError, so codec pipelines over files/sockets don't work with Stream.transform.

Architecture fit

• Add per-codec chunked sinks that carry the bit-accumulator state across chunks, as an additive path beside the fast whole-buffer convert() — the whole-buffer fast path must not regress (differential + benchmark proof).
• Start with the bit-regrouping codecs (base16/32/64) where the accumulator is the generic BitEncoder/BitDecoder state.

Effort: M–L

Checklist

• startChunkedConversion overrides carrying accumulator state
• Tests: chunk-boundary splits at every offset produce identical output
• Differential vs whole-buffer convert() — byte-identical
• Benchmark: whole-buffer path shows no regression (numbers reported)
• README / CHANGELOG / example updated
• Format / analyze / test (vm + node) green