Engine: Fix UTF-8 mojibake (乱码) during streaming

[leo71233826]

Summary

This PR fixes a critical streaming issue where multi-byte UTF-8 characters (e.g., Chinese, Japanese, Emojis) would occasionally render as replacement characters (\ufffd) when a token boundary splits the character's byte sequence.

Motivation

In LLM streaming, a single token might only contain a partial byte sequence of a multi-byte character.

• The Bug: The previous logic used a fixed-length truncation (based on tool-call markers) and decoded the resulting bytes immediately. If the truncation hit the middle of a UTF-8 sequence, it produced a decoding error.
• The Fix: Introduced a trailing-byte validator that detects incomplete UTF-8 sequences at the end of a buffer and holds them back until the next chunk arrives to complete the sequence.

Changes

• Stream.swift: Added a bitwise check on trailing bytes in the performOneGenerationPass loop to ensure we only emit complete UTF-8 sequences.
• Tokenizer.swift: Updated NaiveStreamingDetokenizer to avoid emitting partial segments when \ufffd is detected at the end of the decoded string.

Verification

• Tested with Qwen3.6-35B-A3B on Chinese prompts.
• Confirmed that "mojibake" (random \ufffd inside words) is eliminated during high-speed streaming.
• Verified zero performance regression on ASCII-only text.