selective validation in TensorDomain constructor

[jjsjann123]

Original conversation started from this thread

TensorDomain constructor run domain equivalence check on logical to loop and logical to allocation, which makes sense for most TVs.

Meanwhile, special cases like:

1. ScatterOp where the loop domain doesn't cover the entirety of logical domain, see link;
2. PreprocessGroupedMatmulInputSf has larger allocation domain than logical domain, see link.

These special cases means, when we clone a TensorDomain with some modification, we shouldn't blindly run equivalency checks on domains that's inherited, but rather only on new domain that's being modified.

e.g. In transform_replay.cpp, we have patterns like

 870     TensorDomain* replayed = IrBuilder::createInContainer<TensorDomain>(
 871         consumer->container(),
 872         consumer->getRootDomain(),
 873         consumer->getLogicalDomain(),
 874         consumer->getAllocationDomain(),
 875         new_loop,
 876         consumer->domain()->contiguity());

Where only loop domain is changed, and hence validation check should only be limited to that.

@naoyam proposed to add a TensorDomainBuilder to facilitate such pattern.

auto replayed = TensorDomainBuilder(consumer).loop(new_loop).build();