fix(gpu_test): isolate concurrent test sessions

[tetsuo-cpp]

Summary

• assign each GPU pytest session a unique Vast.ai label and bind it to the instance ID returned by the current SDK
• inventory existing WarpForth instances at startup without deleting resources owned by other runs
• generate a per-session Ed25519 identity and attach only its public key to the owned Vast container, so fresh orbs and CI need only VASTAI_API_KEY
• reconcile every ambiguous create response by exact label and verify teardown after scoped destroy requests
• add mocked concurrency/lifecycle tests, run them in CI, and document parallel-run costs and orphan recovery
• migrate from the legacy vastai-sdk package to the current official vastai package

Verification

• env -u VASTAI_API_KEY uv run pytest -q -m "not gpu" — 8 passed, 37 deselected
• uv run ruff check gpu_test/
• uv run ruff format --check gpu_test/
• cmake --build build --target check-warpforth — 114 passed
• uv lock --check
• git diff --check
• paid Vast.ai smoke test from a fresh orb: two independent pytest processes ran test_addition and test_subtraction concurrently; both attached distinct per-instance SSH keys, reached SSH readiness, compiled the runner, executed their kernels, and passed
• final Vast inventory confirmed every test-created instance was absent and the pre-existing account instance remained untouched

The paid smoke test exposed a Vast response that reported failure after creating a contract; the final follow-up commit handles all responses without an authoritative ID as ambiguous and reconciles by exact label rather than issuing another create request.