fix(providers): buffer partial UTF-8 across SSE chunk boundaries

[kaizhou-lab]

Problem

Streaming CJK (Chinese/Japanese/Korean) model output intermittently dropped characters, replacing them with � (U+FFFD). Reported via Sentry ELECTRON-3NH / ELECTRON-3P5 (OpenCode Zen deepseek-v4-flash-free), but the defect affects all streaming providers on the OpenAI-compatible / Anthropic paths.

Root cause

crates/aion-providers/src/stream_process.rs decoded each raw bytes_stream() chunk independently with String::from_utf8_lossy(&chunk) before handing the text to the &str-accumulating framers. When a multibyte UTF-8 character (a 3-byte CJK char) is split across two network chunks, each half is lossy-decoded separately and each incomplete fragment becomes U+FFFD — so the character is lost. The signature (isolated � exactly at CJK character boundaries, surrounded by byte-perfect text) is a textbook split-multibyte-across-chunk-boundary bug.

Fix

• Add a byte-level Utf8StreamDecoder carry buffer in framing.rs that only emits complete UTF-8 sequences per chunk and retains the incomplete trailing bytes until the next chunk.
• Use it in the OpenAI, OpenAI-Responses and Anthropic SSE paths; flush any remainder at the true end of the stream.
• The Bedrock path already buffers raw bytes and is untouched.

Tests

• cargo test -p aion-providers green (222 + 4 + 9 + 11 passed).
• New framing_test cases: a CJK string (权限管理服务器) split mid-character across chunks reassembles with zero U+FFFD; a legitimately-truncated final byte is handled by flush() without panic; ASCII pass-through unchanged.