ci: overlap CUBRID boot with pip install and cache pip downloads

[yeongseon]

Summary

Attacks the real smoke-test bottleneck. On the last run, container boot +
dependency install was ~80s while the actual example suite ran in ~14s. This PR
targets the 80s fixed cost:

• Overlap container boot with pip install: the CUBRID container now starts
  before dependency install, so its image pull and engine initialization run
  concurrently with pip. The readiness gate afterward usually finds the engine
  already up instead of waiting on a cold start after install finishes.
• Cache pip downloads: ~/.cache/pip is cached with a rolling key, so the
  large stable wheels (pandas, matplotlib, numpy, sqlalchemy) are not
  re-downloaded every run. The key is intentionally not version-hashed — the
  target packages (pycubrid, sqlalchemy-cubrid, cubrid-mcp-server) float to
  their latest release, and pip still queries PyPI for newer versions.

What is unchanged

• Full-suite coverage semantics from ci: scope smoke verify to changed examples on pull requests #38 are untouched (scope step still runs
  first; VERIFY_PATHS logic is identical).
• The readiness gate, external-connectivity check, and make verify are the
  same; only step ordering and a cache layer changed.

Validation

• YAML parses; step order verified: Start CUBRID → Setup Python → Cache pip →
  Install deps → Record versions → Wait readiness → external connect → verify.
• Correctness rests on the readiness gate (unchanged), which still blocks until
  csql SELECT 1 succeeds, so starting the container earlier cannot race the
  examples.