test: re-enable cumsum tests on TensorRT-RTX

[SandSnip3r]

Summary

• remove the blanket TensorRT-RTX skip from the cumsum Dynamo converter tests
• let current-main CI determine whether the TRT-RTX 1.4-era regression tracked by NVBugs 6060938 still reproduces

Local validation

• Windows 11, NVIDIA GeForce RTX 5090
• PyTorch 2.14.0.dev20260721+cu130
• Torch-TensorRT current main (2100545)
• TensorRT-RTX 1.6.1.13 from current local TRT/Myelin builds
• static and dynamic cumsum cases pass with immutable_weights=False and require_full_compilation=True
• historical pure-TensorRT refit/INT64 loop reduction also passes