Fix FP8 current scaling FakeTensor output shape

[sanjana658]

Description

Please include a brief summary of the changes, relevant motivation and context.

Fixes # (issue)

Summary

• Return a scalar inverse-scale tensor from the fake implementation of tex::fp8_cs_quantize, matching the eager output metadata.
• Add a regression test using torch.library.opcheck with test_faketensor.

Validation

• Python syntax compilation passed for both changed files.
• git diff --check passed.
• Native CUDA opcheck test has not been run locally because the local environment has no CUDA GPU and TransformerEngine's compiled package is not installed.

Related issue: #3636

[greptile-apps]

Confidence Score: 5/5

[Medium impact] The PR appears safe to merge; the fake inverse scale now matches the eager output.

Summary

Makes the fake output of tex::fp8_cs_quantize return a scalar inverse scale, matching eager execution.

• Adds a focused torch.library.opcheck regression test.
• Adds the test to the L0 PyTorch runner, addressing the earlier CI coverage comment.
• No new actionable issues were found. CUDA execution was not verified during this review.

_{Reviews (3) · Last reviewed commit: "test: include FP8 fake tensor opcheck in..." · Reviewed by Greptile}

[sanjana658]

Hi maintainers, I've addressed the CI coverage feedback by adding the regression test to the L0 PyTorch test runner, and the DCO check is now passing. The PR is ready for review. Could someone please approve the pending workflows and review the changes when convenient? Thank you!

[ptrendx]

Please add this test to the existing test_onnx_export.py instead rather than creating a completely new file.