fix(dynamo): handle symbolic converter parameters

[narendasan]

Summary

• inspect FX nodes nested in list, tuple, and dict converter arguments when validating dynamic shapes
• make runtime split sections decline TensorRT conversion and fall back cleanly instead of reaching Python offset arithmetic as ITensors
• normalize dynamic arange start, end, and step operands to one TensorRT Fill dtype
• retain the established Int64 output contract while handling rank-0 LINSPACE alpha inputs

Issues

Closes #4606
Closes #4625

Tests

• pytest tests/py/dynamo/conversion/test_split_aten.py tests/py/dynamo/conversion/test_arange_aten.py -n0 (30 passed)
• original reproducers for 🐛 [Bug] split validator does not inspect list-valued args, so runtime section sizes abort compilation #4606 and 🐛 [Bug] arange passes a rank-1 tensor to LINSPACE's rank-0 alpha input; the API error is logged, not raised #4625 no longer reproduce; both symbolic-start and static-start arange models compile and match eager
• syntax, Ruff, Black, and whitespace checks pass

Stack

7 of 7. Base: #4632.