fix: preserve binary multipart chunks

[Guflly]

Summary

• normalize multipart enumerator chunks to binary before buffering
• keep bounded and whole-stream reads in binary encoding
• add regression tests for mixed UTF-8 and binary chunks

Fixes #317.

Validation

• focused adapter tests pass
• RuboCop passes
• the full util test file has one existing Windows CGI filename failure, reproduced on current main

[jbeckwith-oai]

Reviewed the adapter and all production call sites. Converting enumerated chunks to ASCII-8BIT before buffering preserves the exact bytes, eliminates UTF-8/BINARY compatibility failures, and leaves IO/StringIO behavior and transport framing unchanged. Bounded reads remain streaming with memory limited to the current source chunk/buffer; the unbounded-read path retains its existing whole-stream semantics. Regression coverage exercises both bounded and whole-stream mixed-encoding reads. Local validation: util suite 38/38 (160 assertions), focused mixed-encoding tests, an end-to-end multipart text+binary byte-integrity probe, RuboCop on both changed files, and git diff --check all pass.