except exatn::sync(TensorOperation) which is always local.
Bug fix in replicateTensor.

Exposed a couple of ExaTN APIs to Python

- Merge tensors API that returns pattern string.

- Create tensor (in num server) from an existing tensor.

Also, added a simple example.

(This example was adapted from discussion with Andrew Chen)
Signed-off-by: Thien Nguyen <nguyentm@ornl.gov>

TensorOpTransform/Slice/Insert need the original tensor shape ...

One problem is still there: Output tensors are moved to accelerators by default
but currently they are not discarded efficiently when TRY_LATER occurs.
Need some kind of LRE active cache eviction policy ...

Fixed round-off issues for FP32 full contractions in CP-TAL.