fix(dynamo): handle dynamic scatter and index updates

[narendasan]

Summary

• skip scatter_add decomposition when its extent is data-dependent so supported dynamic graphs remain compilable
• normalize index_put update dtype to the destination tensor before TensorRT scatter
• represent None-index expansion with dynamic arange/broadcast logic instead of one index per statically known element
• make static arange conversion safe under FakeTensorMode and clamp inferred dynamic profile minima

Issues

Closes #4602
Closes #4609
Closes #4621

Tests

• pytest tests/py/dynamo/conversion/test_index_put_aten.py -n0 (74 passed)
• pytest tests/py/dynamo/lowering/test_decompositions.py tests/py/dynamo/models/test_dynamic_shape_user_bounds.py -n0 (94 passed)
• original reproducers for 🐛 [Bug] scatter_add decomposition requires a static extent, aborting on a data-dependent one #4602, 🐛 [Bug] index_put casts scatter indices but not updates, so IScatterLayer rejects mismatched types #4609, and 🐛 [Bug] index_put with a None index materialises one index per element, failing on a runtime-known extent #4621 no longer reproduce

Stack

5 of 7. Base: #4630. Follow-up: narendasan/fix-converter-profile-edge-cases.