fix(ollama): soft-fail on empty response in generate_from_raw

[planetf1]

Description

Fixes #599

generate_from_raw was intermittently returning empty-string results. After investigation (see #16326 which we have now closed), the root cause was in the mellea client: asyncio.gather(..., return_exceptions=True) silently converted any exception — timeout, ResponseError, connection reset while the model was loading — into ModelOutputThunk(value=""). This was fixed separately in #1163 (dropped return_exceptions=True).

This PR adds a second, belt-and-braces layer for the rare but real case where Ollama genuinely returns HTTP 200 with response: "" and done: True (e.g. EOS sampled as the first token — runner.go:546–552 in Ollama, where a TODO comment acknowledges the issue). In 4400 requests across warm and cold-load conditions we never observed this path, but the code path exists and is worth handling explicitly.

mellea/backends/ollama.py — adds an elif response.done and not response.response and not response.thinking: branch. Logs a warning and attaches a RuntimeError and the raw GenerateResponse to generate_log.extra for operator inspection. The and not response.thinking guard avoids false-positives on thinking models that return an empty .response alongside a non-empty .thinking field. Does not re-introduce return_exceptions=True — that removal from #1163 is preserved.

test/backends/test_ollama.py — adds a module-scoped _ensure_model_warm autouse fixture so that running this file in isolation (outside the full conftest warm-up path) doesn't start cold. Removes the @pytest.mark.xfail from test_generate_from_raw — the original failure was the return_exceptions=True bug, now fixed; the test has passed consistently across 1000+ trials.

test/backends/test_ollama_unit.py — unit tests for the new branch (no live Ollama required): empty done-response soft-fails with RuntimeError attached; thinking-model response with response="" is not flagged; one empty slot in a batch of three doesn't discard the other two results.

Testing

# Unit tests — no Ollama server required
uv run pytest test/backends/test_ollama_unit.py -v
# 19 passed

# 1000-trial soak: 4000 requests, 0 empty responses
uv run --with httpx python /tmp/hammer1000.py 1000
# DONE: 1000 trials, 4000 requests, 0 empty responses, 398s elapsed

# Verify xfail removal — test now passes clean
uv run pytest test/backends/test_ollama.py::test_generate_from_raw -v
# 1 passed