#572 serve: return unfinished tool call text as content

Beschreibung

Tool-enabled API clients can get a clean, empty completion when the model ends its turn inside an announced call. In the HTTP reproduction with the mock engine on v0.1.38, the Chat Completions and Messages endpoints returned empty content for JSON and SSE with thinking on or off.

This builds on the code @beatcollector posted in [issue #537](https://github.com/Niko1221/Strata/issues/537#issuecomment-5961944986). The parser now returns the raw buffered unfinished call text as content, without closing the JSON arguments or emitting a completed tool_call. The existing complete-call tests still pass. The parser regression follows the posted chunks/feed/finish reproduction with actual function and parameter openers, plus THINK_END when thinking is enabled.

`.venv/bin/python -m unittest serve.test_server -q` passed (120 tests). The full serve suite passed (199 tests, 7 skipped). The full 30-module Python suite ran 486 tests with the same 47 pre-existing setup failure testcase/subcase labels, in the same order, as both 484-test baselines; each had 7 skips. The new regression was red in 2 tests with 8 failing subcases, and the full UnfinishedToolCall class passed (7 tests). Reverting the source reproduced the same 8 failures with the MD5 mutation confirmed. Native GPU/model tests were not run.

Co-authored-by: beatcollector <[email protected]>