fix(dynamo): normalize scalar values at engine boundaries

[narendasan]

Summary

• materialize scalar partition inputs as zero-dimensional CUDA tensors before calling the Tensor[] execute_engine schema
• restore scalar engine outputs with aten.item for their original consumers
• skip scalar input shape metadata and select output placement from the first tensor input
• keep inlined execute_engine graphs compatible with AOTInductor serialization

Issues

Closes #4614
Closes #4616
Closes #4617
Closes #4620

Tests

• pytest tests/py/dynamo/models/test_exporter_inlining.py -n0 (7 passed)
• pytest tests/py/dynamo/models/test_meta_kernel_shape_inference.py -n0 (12 passed)
• original reproducers for 🐛 [Bug] a SymInt engine output is returned as a 0-d tensor and never converted back for scalar consumers #4614, 🐛 [Bug] engine meta kernel reads .shape on scalar inputs despite recording is_scalar: True #4616, 🐛 [Bug] engine meta kernel takes the output device from inputs[0], failing when the first input is a scalar #4617, and 🐛 [Bug] inlined execute_engine gets a SymInt where its schema declares Tensor[], breaking AOTI serialization #4620 no longer reproduce
• 🐛 [Bug] inlined execute_engine gets a SymInt where its schema declares Tensor[], breaking AOTI serialization #4620 AOTInductor package creation succeeds

Stack

2 of 7. Base: #4627. Follow-up: narendasan/fix-device-placement.