ci(python): GraalPy's unit tier stops on an xdist worker MemoryError under the 2 GiB per-process heap #1753

Describe the bug
Since #1749, the Python graalpy-25.4 (experimental) leg no longer loses its runner. It now ends early on its own.

On #1749's run (job 116172356218), an xdist worker raised MemoryError inside pytest_runtestloop. The controller then stopped the session with INTERNALERROR, and pytest exited 3 after 986 s, at 34 failed, 569 passed, 4500 subtests passed.

The runner itself had memory to spare. The last sample showed 8069 of 15988 MiB used and 7919 MiB available, inside the 12G scope, with 0 OOM kills.

Expected behavior
The GraalPy leg runs the unit tier to completion, or to its 110-minute timeout, and reports real test results. It should not stop on a worker running out of memory.

Additional context

• The leading guess, not yet measured, is the 2 GiB heap that every GraalPy process gets from GRAAL_PYTHON_VM_ARGS=--vm.Xmx2g (experimental.yml, the GraalPy containment comment). Each of the 3 xdist workers runs under that cap, so one worker can exhaust its own heap while the machine still has half its RAM free.
• The 34 failures before the abort are a separate question: they may be GraalPy incompatibilities rather than defects.
• Follows ci(python): the GraalPy experimental leg loses its runner 18–25 minutes into its tests on every recent run #1740, whose expected behaviour (the leg ends on its own result, not on a lost runner) is met.