Fix flaky test_siblings_without_eval - #3621

)

compile_dfs deep-copies the traced graph so the cache owns a private mutable
copy, then redirects every consumer to the copy and leaves the originals to be
freed by reference counting. Single-output originals are. Multi-output
originals are not: each output holds its siblings in ArrayDesc::siblings, so
the group refers to itself. array::~array() breaks that cycle, but the
redirection goes through array::operator=, which only reseats array_desc_ and
never runs the destructor, and the vector teardown happens while the
not-yet-remapped parents map still holds copies so the destructor's count test
fails. The orphaned group then keeps its whole input cone alive -- including
any arrays the compiled function captured -- for the lifetime of the process.

Trace nodes hold no buffers, so the visible cost is the captured constants:
65,536 B per re-trace in the issue's repro. The trigger is a new trace (fresh
closure identity, or new shapes), not an eviction, which is why a plain
function is unaffected -- its stable fun_id means it traces once.

Detaching the replaced originals where they are created also fixes
mx.export_function, the only other caller of compile_dfs, which leaks
identically with no compile cache involved at all.

The guard fires only for nodes actually replaced: graph inputs, graph outputs
and primitive-less nodes all map to themselves. Sibling groups are
all-or-nothing because io_set is seeded with o.siblings(), so
`a, b = split(x); del a; eval(b)` never enters the branch. The Primitive
survives because the copies co-own it via primitive_ptr().

Regression test uses the std::weak_ptr<array::Data> probe from ml-explore#1590/ml-explore#3621
rather than an RSS or get_active_memory assertion. Verified CPU-only
(MLX_BUILD_METAL=OFF): the test fails with this hunk removed and passes with
it, and the full C++ suite is 245/245 cases / 3240/3240 assertions.