Rewriter: don't crash fusing Relu/Clip chains when the Clip input has no type

[Kayvan Zahiri (Kayvan-Zahiri)]

rewriter.rewrite() raises on a valid model when the input of a fused Clip has no type information, for example when shape inference hasn't been run:

x1 = Abs(X)
x2 = Clip(x1, min, max)
Y  = Clip(x2, min2, max2)

AttributeError: 'NoneType' object has no attribute 'numpy'

extract_min_max read the dtype from node.inputs[0].dtype.numpy(), and x1.dtype is None here. All three Clip rules (Clip/Clip, Clip/Relu, Relu/Clip) hit it, whether the bounds are initializers or Constant nodes. optimizer.optimize() doesn't, because it infers types first.

The spec gives min and max the same type as the input, so the dtype now comes from the bounds when the input has none. The one case left without a dtype is Clip(Relu(x)) with no bounds and an untyped x. There, check() now fails the match instead of building a float64 min.

Tests: a new class runs the three patterns on an untyped intermediate, without shape inference, and checks the no-bounds failure. All 4 tests fail on main with the error above. Rewriter tests: 414 passed. test_fuse_pad_into_conv_6/_7 fail on main too in my environment.

Found while fuzzing the rewrite rules on untyped models, with Claude Code's help.