fix(connector): preserve suffix stores on gpu prefix hits

[GentleCold]

Summary

• Plan missing rolling-prefix suffix stores before returning zero external tokens.
• Preserve existing behavior that skips already cached DaseR slots.
• Extend the prefix-hit regression test to cover a longer GPU-local hit.

Motivation

When vLLM already has a longer GPU-local prefix than DaseR, the DaseR hit contributes no external tokens. The scheduler returned before preparing the missing suffix store, leaving the lower-tier prefix permanently shorter until GPU eviction forced recomputation.

Test plan

• Unit tests added / updated
• PYTHONHASHSEED=0 pytest -q -m "not integration" --ignore=tests/integration --ignore=tests/connector/test_daser_connector.py --ignore=tests/connector/test_gds_transfer.py tests/ passes: 353 passed
• Connector and server tests pass: 240 passed
• Integration tests: N/A; scheduler lifecycle behavior is covered by the connector regression
• Benchmark / e2e smoke: validated separately with TP2 long-context replay; no benchmark code is included
• pre-commit run --all-files passes

Checklist

• pre-commit run --all-files passes
• Type hints and docstrings on all new/modified functions
• New features and bug fixes include tests
• No regressions - existing tests still pass