Add fx passes to support unbounded dynamism in torch op arg

[lsy323]

Add FX passes to support dynamism on torch ops that take symbolic dim size as parameters (e.g. `aten.view(x, [sym_size, -1]). The FX passes groups the generation of symbolic dim size and the torch op into a new XLA op. Grouping them together is necessary to lowering these ops with dynamism, because the operations on symbolic size cannot be traced in LTC (context in #6393). Once the source of the symbolic size and the consuming torch ops are captured in a single XLA op, it becomes feasible to lowered to HLO/StableHLO with dynamism semantics.

The FX passes will run automatically if the exported program has symbolic shape input.

The following ops are fused into XLA ops:

• sym_size.int + aten.expand => xla.dynamic_expand
• sym_size.int + (mul) + aten.view => xla.dynamic_view

Some torch ops are generating aten.view with symbolic dim size during decomposition in upstream PyTorch, or existing lowering logic in torch_xla. FX passes are introduce to handle these ops as well.

• aten.native_layer_norm
• aten.group_norm
• aten.select
• aten.unsqueeze

Other changes:

• Adds dynamism support to rsub
• Adds dynamism support to mean and var
• Move torch_xla/test/stablehlo/utils.py to torch_xla/utils/stablehlo_test_utils.py, to fix path not found issue when running tests with pytest.
• Added unbounded dynamism unit tests for more ops.

Test:

• Added test script for each of the FX passes, and unbounded dynamism unit tests for each supported op. Correctness of the lowering is done by comparing torch output and tf.saved_model inference result.