DID-parallelize a loop split in Python.

[wujingyue]

For #2563

List of major changes:

1. Add FusionDefinition.sched.set_allocation_as_loop to set the allocation domain of a TensorView to be the same as loop.
2. Add a convenience helper shard_tensor to be used in unit tests. This leads to some refactoring on the C++ side.
3. Remove _create_device_mesh with a DeviceMesh constructor. This way, mesh construction doesn't require self.sched and is made more flexible.
4. Add setup/finalizeMultideviceSchedule to set the active fusion for multidevice_schedule.
5. Add a Python test to exercise DID parallelization of loop domains.

[wujingyue]

!test

[jjsjann123]

lgtm. Failing tests looks like something with CI machines? I restarted the tests 🤞

[wujingyue]

lgtm. Failing tests looks like something with CI machines? I restarted the tests 🤞

H100 tests have been failing in CI for quite a while. I already tagged @xwang233 in #3284 (comment).

[wujingyue]

!test

[wujingyue]

!test