test(runtime): dead_owner_side_tables races with itself — 15 failures in 200 filtered runs on main #7975

Summary

gc::tests::dead_owner_side_tables has its own intra-module race, in the same
family as #7946/#7955: two of its cases fail ~7 % of runs when the module's
tests are co-scheduled, and the failure is present on origin/main and is NOT
caused by #7974.

Reproduce (release build of the perry-runtime lib test binary):

perry_runtime --test-threads=10 dead_owner_side_tables

i.e. unchanged by #7974 — filed so nobody mis-attributes it to that PR.

Failing cases and messages:

• test_dead_arguments_object_entry_pruned_on_full_gc
  (dead_owner_side_tables.rs:488) — the common one:

  dead arguments object's ARGUMENTS_OBJECTS entry must be pruned
  (one insert per call of any function referencing `arguments`)
  

• test_dead_owner_descriptor_entries_pruned_on_full_gc
  (dead_owner_side_tables.rs:280), less often.

Why it does not show at CI's parallelism

Both arms of a full-suite run are clean: 300 runs of the unfiltered binary
at libtest's default parallelism gave 0 failures on either binary. libtest
spreads a 2 200-test suite out enough that these particular cases rarely land
together. Filtering to the module packs them onto 10 threads and the rate jumps
to ~7 %. So this is currently latent rather than merge-blocking — but it is the
same defect class, and anyone bisecting with a --filter will hit it and
believe their change caused it.

Likely shape (not confirmed)

Both cases assert that an owner with NO roots is dead at a full collection, and
both take GcTestIsolationGuard. The other tests in the module run collections
and install/clear address-keyed side-table entries concurrently
(ARGUMENTS_OBJECTS, PROPERTY_DESCRIPTORS, ACCESSOR_DESCRIPTORS), so the
suspects are the usual two: a concurrent collection on another thread
completing the prune before/after this thread's full_gc(), or an address
recycled into a sibling test's allocation between the plant and the assert.

The #7954/#7974 remedy applies if it holds: the prune algebra is a function of
(side table, valid-pointer set), so it can be driven on tables the test owns,
with the shipped registration pinned read-only.

Acceptance

≥ 200 consecutive --test-threads=10 dead_owner_side_tables runs green, with
the full suite at default parallelism unchanged.