Fix byte-limited reads for Enumerator-backed request bodies

[fallintoplace]

Summary

• make Enumerator-backed ReadIOAdapter#read(max_len) return at most max_len bytes
• use byte-oriented buffering so multibyte strings follow IO#read semantics
• preserve buffered bytes when an unbounded read follows a partial read
• match IO behavior for zero-length reads, EOF, negative lengths, and out_string
• add regression coverage for bounded reads, multibyte data, buffering, and output buffers

Why

This adapter feeds streamed request bodies into IO consumers such as IO.copy_stream. Returning more bytes than requested can break those consumers and corrupt multipart upload boundaries; character-oriented slicing also makes the result depend on string encoding instead of the requested byte count.

Fixes #274.

Test plan

• bundle exec rake TEST=test/openai/internal/util_test.rb test
• bundle exec rake lint

[jbeckwith-oai]

Thank you so much for the careful and comprehensive work here, @fallintoplace. Your analysis of byte-oriented max_len reads, preserving buffered data across bounded and unbounded reads, EOF and out_string semantics, and negative lengths became the backbone of the canonical fix.

Since this PR was opened, ReadIOAdapter moved into its SDK-owned file and main advanced substantially. We carried your work forward in #354, rebased it onto current main, combined it with the mixed-encoding fix for #317, preserved native IO/StringIO delegation, and added end-to-end and randomized contract validation. That replacement is now fully green across the Ruby matrix, lint, signatures, packaging, and CodeQL.

I’m closing this PR only because #354 supersedes it—not because the contribution wasn’t valuable. It materially shaped the final implementation and gave us the complete contract we needed. Really appreciate the time, rigor, and care you put into this. Thank you! 🙏