Add ndtri_exp scalar Op

Fix rewrites that were relaxed from Advanced(Inc)Subtensor1 Op but re…

…lied on their restrictions

AdvancedSubtensor1 / AdvancedIncSubtensor1 guaranteed a single 1-D
integer
index on the leading axis. Three rewrites retargeted onto the general
Ops
still relied on that; the last two returned wrong results silently.

- local_useless_AdvancedSubtensor1 gated on idx_list == (0,), which a
boolean
  mask also satisfies, and crashed comparing it against arange. Deleted:
  local_adv_idx_to_slice already covers the arange take.
- local_set_to_inc_subtensor ignored idx_list and compared only the
first
  index, so a read of x[:, i] matched a write to x[i] and the collapsed
inc
  hit the wrong axis. It now compares and rebuilds with all of them,
gated on
  _advanced_indices_jointly_unique.
- local_subtensor_make_vector mapped a constant index through int(),
reading a
  boolean mask as the positions 0/1. Left for bool_idx_to_nonzero
instead.

Patch numba get_item_pointer2 for A layouts

Move _compile_mode responsibility to mode.optimizer

Don't bail numba core-shape rewrite on RVs that already feed the node

introduce_explicit_core_shape_rv refused to wrap a RandomVariable when its
core shape read the shape of another RandomVariable already feeding it (e.g.
an MvNormal whose random `mean`'s trailing dimension is the core shape). That
RV is computed before the node, so reading its shape needs no extra draw --
but `applys_between(node.inputs, [core_shape])` still yields such an input RV
as the owner of a boundary variable, so the node was left unwrapped and later
crashed numba funcify with "necessary to replace RandomVariable with
RandomVariableWithCoreShape".

Block the ancestor walk at the node's own inputs and only bail on a
RandomVariable reachable past them (a genuine extra draw for the shape).

Deprecate the clone_inner_graph(s) machinery

Cloning an outer graph used to optionally deep-clone the inner graphs of
HasInnerGraph ops (clone_inner_graph(s)=True). Now that every inner-graph op
(Scan/OpFromGraph/Composite/ScipyWrapperOp) is immutable -- clone() returns self --
that deep-clone is always a no-op, so drop the branch throughout the clone machinery
(Apply.clone, Apply.clone_with_new_inputs, clone, clone_get_equiv,
FunctionGraph.clone, rebuild_collect_shared). The public clone_inner_graph(s) kwarg is
kept as a deprecated, ignored no-op that emits a FutureWarning; only the internal
clone_node_and_cache (and its op-clone caching) loses it outright.

Don't register indexed-elemwise fusion by default

The indexed-elemwise fusion (IndexedElemwise), introduced in 3.0.4, can
trigger a RecursionError during OpFromGraph shape inference on some numba
graphs (e.g. FullRankADVI). It was too risky to ship in a minor release;
disable its default registration on the 3.0.x line until the proper fix
(non-recursive core-shape construction) is backported from main.

Avoid deep recursion in OpFromGraph.infer_shape

clone_replace rebuilds the shape graph through the recursive
clone_v_get_shared_updates, which exceeds Python's recursion limit on
deep inner graphs (e.g. the fused IndexedElemwise Composites produced by
numba-mode FullRankADVI), raising RecursionError during infer_shape. Use
the iterative graph_replace instead.

Revert "Remove custom slice hashing workaround for Python <3.12"

This reverts commit 4ed3aff.

Fix local_subtensor_of_batch_dims when encountering stale broadcast t…

…ypes (#2167)

Co-authored-by: Ricardo Vieira <ricardo.vieira1994@gmail.com>