server: preserve reasoning_content token boundaries

[rareba]

Summary

• keep OpenAI reasoning_content SSE deltas aligned to decoded-token boundaries
• retain one-token look-behind while checking for the closing think marker
• replace the regression that expected arbitrary byte splitting

Root cause

The thinking stream held a fixed seven-byte suffix to detect a possibly split
closing think marker. Every update therefore emitted the current decoded token
minus seven bytes, then emitted that tail with the next token.

Remembering the previous decoded-token end preserves whole-token deltas while
keeping the close-tag and UTF-8 safeguards intact. The only cost is one token
of reasoning-stream latency.

Impact

Clients that process individual reasoning deltas now receive stable token
units instead of character fragments. Reassembled output is unchanged.

Validation

Windows 11 host, WSL Linux CPU/reference build; no model required:

• focused server test binary: ds4 tests: ok, server: OK
• make -j2 cpu

Closes #685

[Flor1an-B]

Nice diagnosis and fix. Since your diff is scoped to the OpenAI path, opened #712 to apply the exact same reasoning_token_end technique to responses_sse_stream_update() and anthropic_sse_stream_update() — the other two places #685 pointed at with the identical shape. Not touching your diff at all, just extending the same idea to the rest of the file.