fix(llm): avoid double execution in schedule_stream_flow

[Muawiya-contact]

Summary

Fixes a control-flow fall-through in Scheduler.schedule_stream_flow that caused
the flow to be built, run, and streamed twice when no reusable pipeline
exists (manager.fetch() returns None).

Closes #360.

Root cause

The if pipeline is None: branch builds a fresh pipeline, runs it, streams the
result via post_deal_stream, and caches it with manager.add(pipeline) — but
it was missing a return at the end. Execution therefore fell straight through
into the reuse try block below, where the just-built (now non-None) pipeline
was prepared and run a second time.

Impact: duplicate pipeline execution, extra LLM calls, and doubled streaming
output on the first call for a given flow.

The synchronous counterpart schedule_flow does not have this bug because its
if pipeline is None: branch ends with an explicit return res.

Fix

Add a bare return after manager.add(pipeline) in the if pipeline is None:
branch, so the build/run/stream path terminates instead of falling through.

A bare return is used (not return res) because schedule_stream_flow is an
async generator, where return <value> is not allowed.

Tests

Adds hugegraph-llm/src/tests/flows/test_scheduler.py (Layer A / unit marker,
no Docker, network, or real LLM):

• test_stream_flow_runs_once_when_no_reusable_pipeline — forces
  manager.fetch() to return None and asserts the flow is built, initialized,
  run, and streamed exactly once, the stream output is not duplicated,
  manager.add is called once, and the reuse/release path (prepare /
  release) is never entered. This test fails on the previous code and passes
  with the fix.
• test_stream_flow_rejects_unknown_flow_name — covers the ValueError guard
  for unsupported flow names.

The Scheduler is instantiated via __new__ with a single mocked pool entry to
avoid eagerly constructing every real flow/manager in __init__.

How to verify locally

uv sync --extra llm --extra dev
uv run pytest hugegraph-llm/src/tests/flows/test_scheduler.py -v
uv run ruff format --check .
uv run ruff check .

[Muawiya-contact]

@imbajin Could you please review that one?

[imbajin]

Blocking: no. Summary: No obvious issues found in the current head. Evidence: GitHub checks passed; git diff --check refs/remotes/apache/main...HEAD; uv run --extra llm --extra dev pytest hugegraph-llm/src/tests/flows/test_scheduler.py -q; ruff format/check on touched files.