Reuse kompiled contracts.k and foundry.k in Kontrol performance test

[geo2a]

This PR attempts to reduce the influence of the K compiler processes on the Kontrol performance test. To achieve that, we compile everything in-advance, ensuring that there will be no calls to kompile when running the actual proof test:

• run the single test_foundy_kompile test, which is essentially just a wrapper for the foundry pytest fixture that generates and compiles the contracts.k and foundry.k files. This produces the foundry/out/kompiled directory
• when running the actual proof tests, pass the generated foundry directory to them explicitly. This makes sure that the pytest fixture will not be forced to recompile stuff in every pytest-xdist worker.
• Re-using the foundry directory has a side-effect of also reusing the foundry/out/proofs directory. To make sure that we actually run the proofs in master_shell, and not re-use the ones produced by feature_shell, we delete them between the steps.

Some things I've noticed that may be worth accounting for, but not in this PR:

• the server fixture is module-scoped. That means that all tests in a certain test module will reuse the same server instance, individually in each test worker.
  • the corollary of the above is that the proofs that are run later will use a more "tired" server, since the server state will bloat due to a bunch of duplicate (and likely empty) modules. It is not clear if this is a performance issue, but we have certainly not considered it earlier.
• another thing is that the decision on which test to skip is taken dynamically at the start of the test, and the tests, even if skipped, need the fixtures to be ready for them. In most cases, however, the server connection will be reused by other tests in the same module.

[geo2a]

A test run with 7 workers. Still noisy, but perhaps slightly less so? BMCLoopsTest.test_countdown_concrete() are ran twice by the pytest test suite, and hence we're seeing a the cached result. The change to fix this had been merged into Kontrol.

Test	georgy-kontrol-pref-reuse-artifacts time	master-88ef647ec time	(georgy-kontrol-pref-reuse-artifacts/master-88ef647ec) time
Identity.identity(uint256)	23.8	29.13	0.8170271198077583
kontrol/src/tests/integration/test_foundry_prove.py::test_foundry_resume_proof	102.91	118.84	0.8659542241669471
ExpectRevertTest.testFail_expectRevert_bytes4()	159.03	183.62	0.8660821261300512
ExpectRevertTest.testFail_expectRevert_false()	134.99	151.12	0.8932636315510852
ConstructorTest.test_constructor()	166.28	182.09	0.9131748036685156
SymbolicStorageTest.testEmptyInitialStorage(uint256)	61.08	65.53	0.9320921715244925
MockCallTest.testSelectorMockCall()	76.41	79.55	0.9605279698302954
AddConst.applyOp(uint256)	29.27	30.42	0.9621959237343852
SetUpDeployTest.test_extcodesize()	113.52	117.67	0.9647318772839296
ArithmeticCallTest.test_double_add_double_sub(uint256,uint256)	448.49	432.36	1.0373068739013784
BytesTypeTest.test_bytes32(bytes32)	39.13	37.57	1.041522491349481
kontrol/src/tests/integration/test_foundry_prove.py::test_foundry_auto_abstraction	96.22	91.04	1.0568980667838312
Identity.applyOp(uint256)	51.74	47.78	1.0828798660527417
ConstructorTest.testFail_constructor()	118.87	103.62	1.1471723605481567
ArithmeticContract.add(uint256,uint256)	35.82	29.94	1.1963927855711423
ArithmeticCallTest.test_double_add_sub_external(uint256,uint256,uint256)	393.63	303.5	1.2969686985172981
kontrol/src/tests/integration/test_foundry_prove.py::test_foundry_remove_node	46.7	32.43	1.4400246685168054
ExpectRevertTest.testFail_ExpectRevert_failAndSuccess()	197.33	111.55	1.7689825190497537
BMCLoopsTest.test_countdown_concrete()	38.34	0.53	72.33962264150944
TOTAL	2333.56	2148.29	1.0862406844513544

[jberthold]

[goodlyrottenapple]

Looks good. I worry that the more assumptions/fixes we add in our script the more we will have to keep patching it when the CLI arguments/foundry folders/etc. change. I guess time will tell and perhaps these can be up-streamed to kontrol directly.

[ehildenb]

@geo2a this seems like something that we could have made a change directly to Kontrol for. Consider doing so there! (as I think @goodlyrottenapple also suggested).