[test] Disable flaky leakcheck tests, as done in ROOT's cppyy

[guitargeek]

The leak checks in test_leakcheck.py compare RSS before and after a tight call loop, which is unreliable in the presence of arenas and free-lists. The same tests are disabled in ROOT's copy of the cppyy test suite (bindings/pyroot/cppyy/cppyy/test/test_leakcheck.py), so mark them as skipped here with the same reason.

This replaces the macOS-specific xfail marks on test01-test03 and test08, which the unconditional skip supersedes, and with them the now unused IS_MAC and IS_MAC_X86 imports.

test09_initializer_list_argument has no counterpart upstream, as it was added along with the InitializerListConverter fix in #54. It is skipped too: it turned out to be just as sporadic.

[guitargeek]

Coincidentally, test09_initializer_list_argument just showed to be flaky in the CI as well:

=================================== FAILURES ===================================
________________ TestLEAKCHECK.test09_initializer_list_argument ________________

self = <test_leakcheck.TestLEAKCHECK object at 0x10bc79f50>

    def test09_initializer_list_argument(self):
        """Leak check of passing a list as an std::initializer_list argument"""
    
        import cppjit
    
        cppjit.cppdef("""\
        namespace LeakCheck {
            int sum_il(std::initializer_list<int> l) {
                int s = 0;
                for (auto i : l) s += i;
                return s;
            }
        }""")
    
        ns = cppjit.gbl.LeakCheck
    
>       self.check_func(ns, "sum_il", [1, 2, 3])

test_leakcheck.py:300: 
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 

self = <test_leakcheck.TestLEAKCHECK object at 0x10bc79f50>
scope = <namespace cppjit.gbl.LeakCheck at 0x7f91a6259e10>, func = 'sum_il'
args = ([1, 2, 3],), kwds = {}

    def check_func(self, scope, func, *args, **kwds):
        """Leak-check 'func', given args and kwds"""
    
        import gc
    
        # if tmpl_args is provided as a keyword, then this is a templated
        # function that is to be found on each call python-side
        tmpl_args = kwds.pop("tmpl_args", None)
    
        # warmup function
        gc.collect()
        if tmpl_args is None:
            getattr(scope, func)(*args, **kwds)
        else:
            getattr(scope, func)[tmpl_args](*args, **kwds)
    
        # number of iterations
        N = 100000
    
        # The use of arena's, free-lists, etc. means that checking rss remains
        # unreliable, unless looking for consistent jumps, so the leak check will
        # be run M times and only considered failed if it "leaks" every time. In
        # actual practice, the number of fails is 0, 1, or M. Note that the total
        # number of gc objects tracked is always required to remain the same.
        M = 3
    
        # leak check
        fail = 0
        for i in range(M):
            gc.collect()
            pre = len(gc.get_objects())
            last = self.process.memory_info().rss
    
            if tmpl_args is None:
                self.runit(N, scope, func, *args, **kwds)
            else:
                self.runit_template(N, scope, func, tmpl_args, *args, **kwds)
    
            gc.collect()
            assert len(gc.get_objects()) == pre
            if last < self.process.memory_info().rss:
                fail += 1
    
>       assert fail < M
E       assert 3 < 3

test_leakcheck.py:75: AssertionError

I'll disable it too.

[aaronj0]

Coincidentally, test09_initializer_list_argument just showed to be flaky in the CI as well:

Yes this test has failed on the nightlies a couple of times. Although the platform is always intel-mac, I agree that the testing method here is not reliable. There is one more flaky failure when comparing execution times in the numba tests but on very rare occasion.

[aaronj0]

I think all the added skips can be dropped in favour of a single one here

Suggested change

	@mark.skip(reason="disabled due to its sporadic nature, especially fragile on VMs")

[guitargeek]

Coincidentally, test09_initializer_list_argument just showed to be flaky in the CI as well:

Yes this test has failed on the nightlies a couple of times. Although the platform is always intel-mac, I agree that the testing method here is not reliable. There is one more flaky failure when comparing execution times in the numba tests but on very rare occasion.

Good to know! I prefer to keep this PR just about the leakchecks for now, but let's keep the numba tests in mind for later.

[aaronj0]

ruff errors are unrelated, fixed in #84

[aaronj0]

Thanks!