fix: resolve lifted TRT engine custom-object by graph-signature FQN in ExecuTorch export

[shoumikhin]

What's broken

Saving a Torch-TensorRT model with output_format="executorch" fails when the model was compiled with require_full_compilation=False (so some ops stay outside TensorRT) and exported through the standard torch.export path:

RuntimeError: execute_engine node 'execute_engine': placeholder engine
'obj__run_on_acc_0_engine' not found in exp_program.constants

The engine is actually present in exp_program.constants — save() just looks it up by the wrong key.

Why it happens

When torch.export lifts the TensorRT engine (a ScriptObject), it stores it in exp_program.constants keyed by its graph-signature FQN (InputSpec.target) but gives the placeholder node a mangled name (an obj_ prefix). The current code looks it up by the node name/target:

engine_obj = constants.get(engine_node.name) or constants.get(engine_node.target)

Those don't equal the FQN key, so the lookup returns None and the export aborts. The legacy exporter (retrace=False) happened to keep the placeholder name equal to the constants key, so it worked by accident; the modern exporter is the one that breaks.

The fix

Resolve the placeholder via the canonical ExportGraphSignature.inputs_to_lifted_custom_objs mapping (placeholder name -> FQN), then read constants[fqn]. Keep the direct name/target lookup only as a fallback for legacy programs that lack that mapping, and unwrap a FakeScriptObject to its real object so it can be serialized. The same buggy lookup existed in two places (the save serializer in _compile.py and the backend engine-info extractor in executorch/backend.py); both now share one helper in dynamo/_exporter.py.

Repro

import torch
import torch_tensorrt

class M(torch.nn.Module):
    def forward(self, x):
        a = (x + 1.0) * 2.0       # TensorRT
        b = torch.cos(a)          # forced off TensorRT
        return (b * 3.0) + 4.0    # TensorRT

ep = torch.export.export(M().eval().cuda(), (torch.randn(2, 3, 4, 4).cuda(),))
gm = torch_tensorrt.dynamo.compile(
    ep, min_block_size=1, torch_executed_ops={torch.ops.aten.cos.default}
)
torch_tensorrt.save(
    gm,
    "m.pte",
    output_format="executorch",
    arg_inputs=(torch.randn(2, 3, 4, 4).cuda(),),
)

Before: raises the error above. After: writes a .pte containing both the TensorRT and CUDA delegates.

Tests

Adds CPU-only unit tests for the resolver in tests/py/dynamo/executorch/test_api.py (no GPU or ExecuTorch runtime required): signature-FQN resolution, legacy fallback, present-but-incomplete mapping -> None, missing -> None, and FakeScriptObject unwrap.

[cehongwang]

LGTM