Support RandomVariable graphs with scalar shape parameters in JAX backend

[ricardoV94]

This should make it possible to do forward sampling in more PyMC models that use dims to define variables shapes

def test_random_scalar_shape_input():
    dim0 = pt.scalar("dim0", dtype=int)
    dim1 = pt.scalar("dim1", dtype=int)

    out = pt.random.normal(0, 1, size=dim0)
    jax_fn = compile_random_function([dim0], out)
    assert jax_fn(np.array(2)).shape == (2,)
    assert jax_fn(np.array(3)).shape == (3,)

    out = pt.random.normal(0, 1, size=[dim0, dim1])
    jax_fn = compile_random_function([dim0, dim1], out)
    assert jax_fn(np.array(2), np.array(3)).shape == (2, 3)
    assert jax_fn(np.array(4), np.array(5)).shape == (4, 5)

These was already a special rewrite to replace make_vector, expand_dims in the shape of RVs, but without handling these inputs from the outside it wouldn't achieve much for PyTensor users:

pytensor/pytensor/tensor/random/rewriting/jax.py

Lines 38 to 77 in e88117e

	@node_rewriter([RandomVariable])
	def size_parameter_as_tuple(fgraph, node):
	"""Replace `MakeVector` and `DimShuffle` (when used to transform a scalar
	into a 1d vector) when they are found as the input of a `size` or `shape`
	parameter by `JAXShapeTuple` during transpilation.
	The JAX implementations of `MakeVector` and `DimShuffle` always return JAX
	`TracedArrays`, but JAX only accepts concrete values as inputs for the `size`
	or `shape` parameter. When these `Op`s are used to convert scalar or tuple
	inputs, however, we can avoid tracing by making them return a tuple of their
	inputs instead.
	Note that JAX does not accept scalar inputs for the `size` or `shape`
	parameters, and this rewrite also ensures that scalar inputs are turned into
	tuples during transpilation.
	"""
	from pytensor.link.jax.dispatch.shape import JAXShapeTuple
	size_arg = node.inputs[1]
	size_node = size_arg.owner
	if size_node is None:
	return
	if isinstance(size_node.op, JAXShapeTuple):
	return
	if isinstance(size_node.op, MakeVector) or (
	isinstance(size_node.op, DimShuffle)
	and size_node.op.input_ndim == 0
	and size_node.op.new_order == ("x",)
	):
	# Here PyTensor converted a tuple or list to a tensor
	new_size_args = JAXShapeTuple()(*size_node.inputs)
	new_inputs = list(node.inputs)
	new_inputs[1] = new_size_args
	new_node = node.clone_with_new_inputs(new_inputs)
	return new_node.outputs

---

📚 Documentation preview 📚: https://pytensor--1029.org.readthedocs.build/en/1029/

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 82.12%. Comparing base (c1ecbe0) to head (70e3929).
Report is 3 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #1029      +/-   ##
==========================================
- Coverage   82.13%   82.12%   -0.02%     
==========================================
  Files         211      211              
  Lines       49738    49751      +13     
  Branches     8816     8817       +1     
==========================================
+ Hits        40852    40857       +5     
- Misses       6710     6714       +4     
- Partials     2176     2180       +4     

Files with missing lines	Coverage Δ
pytensor/link/jax/linker.py	96.22% <100.00%> (+1.22%)	⬆️

... and 1 file with indirect coverage changes

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[ricardoV94]

This solves pymc-devs/pymc#7348

[lucianopaz]

LGTM. There are just 3 minor changes that I requested before merging.