layout op does not work with tensor metadata

[zasdfgbnm]

In layoutAllocationDomain, the following code:

Fuser/csrc/ops/indexing.cpp

Lines 344 to 362 in 06a21af

	// pad row size: num_groups * (row_multiple - 1) + row_size
	auto pad_to_max_extent = [&](IterDomain* id, int multiple) -> IterDomain* {
	auto* maximum_pad_value_per_group =
	IrBuilder::create<Val>(multiple - 1, DataType::Index);
	// NOTE: we do not use `resize` to represent the padding.
	//
	// resize sounds good in theory, because transformation can propagate across
	// it. In reality, we do not have a protocol to index this operation via the
	// logical to allocation domain transform. I question how much a resize op
	// provides in functionality. More importantly, using resize hits asserts in
	// vectorization analysis (validateDeviceSplit ATM), which doesn't look easy
	// to handle for me.
	Val* padded_ext = SimplifyingIrBuilder::addExpr(
	id->extent(),
	SimplifyingIrBuilder::mulExpr(num_groups, maximum_pad_value_per_group));
	return IterDomainBuilder(id).extent(padded_ext).build();
	};
	alloc_dom.push_back(pad_to_max_extent(logical_dom[0], row_multiple));

We are creating a dangling IterDomain in the allocation domain, which is not connected to the logical domain. However, in tensor_metadata.cpp, especially

Fuser/csrc/tensor_metadata.cpp

Lines 308 to 309 in 06a21af

	ForwardTraverseFromLogicalToAlloc(ee, active_ids).run(tv, logical, alloc);
	BackwardTraverseFromLogicalToAlloc(ee, active_ids).run(tv, logical, alloc);

we rely on the connection between logical domain and allocation domain to infer the shape and stride.