PoC: Convert cond

[dayo09]

For #219

[jinevening]

Redundant file

[jinevening]

Summarized by cline.

Summary of the cond_support change

The diff titled cond_support introduces full support for the higher‑order PyTorch operator torch.cond (conditional execution) throughout the TICO code‑base. The main points are:

Area	What was added / changed
New operator	CircleIf (custom op circle_custom::if_) is registered in tico/utils/register_custom_op.py and a corresponding visitor CircleIfVisitor is added in tico/serialize/operators/op_circle_if.py.
IR mapping	MapSubgraph pass (tico/experimental/controlflow/passes/map_subgraph.py) walks the exported program, extracts the true/false sub‑graphs of a torch.ops.higher_order.cond node, evaluates them with concrete inputs, and creates a Circle if node (torch.ops.circle_custom.if_).
Serialization	circle_serializer.build_circle is extended to handle sub‑graphs: it now iterates over all graph modules (root + sub‑graphs) and exports tensors/operators for each. Helper functions _export_tensors and _export_tensors_for_subgraph were added.
Utility structs	CondArgs and CircleIfArgs dataclasses were added to tico/utils/validate_args_kwargs.py to give a typed view of the arguments of torch.cond and the Circle if op.
Sub‑graph handling	New tico/utils/subgraph.py provides freeze_subgraphs, get_frozen_subgraphs, and get_all_graph_modules utilities to manage and retrieve sub‑graphs during conversion.
Pass infrastructure	PassBase.call signature was changed to accept a graph_module argument, and all passes were updated accordingly. The PassManager.run method now forwards the current graph module.
Shape inference	freeze_subgraphs is called before the main conversion loop so that FakeTensor shape inference works for sub‑graphs.
Other adjustments	Numerous files were updated to use the new graph_module argument, to import the new utilities, and to handle the new if operator (e.g., circle_serializer, convert.py, trace_decorators.py).
Tests / examples	Example scripts (call.py, if.py) and a test module (test/modules/op/cond.py) were added to demonstrate conditional models.

Overall, the change adds a complete pipeline:

1. Detect torch.ops.higher_order.cond in the exported graph.
2. Extract its true/false sub‑graphs, evaluate them with concrete inputs to obtain meta‑values.
3. Freeze sub‑graphs for shape inference.
4. Create a Circle if operator that references the frozen sub‑graphs.
5. Serialize the whole model (root + sub‑graphs) into a Circle binary.

[dayo09]

@jinevening FYI, I refactored the code and removed freeze_subgraph stages just now. It's got simpler. :-D