Do not assume mul-sum pattern inputs are TensorView

[jacobhinkle]

This was a change I made to handle casts that wound up breaking some tests and benchmarks in #2272, leading to dynamic cast errors or segfaults. The solution is to test the type of the left and right hand sides before processing the pattern matching.

[jacobhinkle]

!build

[jacobhinkle]

Multidevice tests are failing. The fusion looks like this

Fuser/tests/cpp/test_multidevice_matmul.cpp

Lines 91 to 108 in b85457e

	TensorView* a = makeContigTensor(3, DataType::Half); // (Mo,Mi,K)
	TensorView* b = makeContigTensor(2, DataType::Half); // (N,K)
	TensorView* a_b = broadcast(a, {false, false, true, false}); // (Mo,Mi,b,K)
	TensorView* b_b = broadcast(b, {true, true, false, false}); // (b,b,N,K)
	TensorView* ab = mul(a_b, b_b); // (Mo,Mi,N,K)
	TensorView* c = sum(ab, {-1}); // (Mo,Mi,N,r)
	fusion->addInput(a);
	fusion->addInput(b);
	fusion->addOutput(c);
	// Sharding M dimension
	auto all_sharded_tvs = {a, a_b, b_b, ab, c};
	for (auto tv : all_sharded_tvs) {
	tv->axis(0)->parallelize(ParallelType::DIDx);
	tv->setDeviceMesh(mesh);
	}
	b->setDeviceMesh(mesh);

This creates an outer M dimension that is parallelized as DIDx. We should ignore such dimensions in pattern matching.

[jacobhinkle]

Merging for now to mostly unblock CI then I will address the multidevice problem in another PR.