[REVIEW] Fix padded dataset creation for aligned device tensors

[Sharon05564]

Description

Fixes cuvsDatasetMakePadded() rejecting device tensors whose row stride already satisfies the required padding alignment.

Previously, an already-aligned device tensor caused the padded-dataset helper to return an error and the Python binding worked around this by inspecting the error message and falling back to a non-owning padded view.

This change:

• adds an opt-in force_copy path to the padded-dataset helpers
• makes the C API use that path so cuvsDatasetMakePadded() always returns an owning copy
• preserves the existing default behavior for C++ callers
• removes the Python error-string fallback
• adds regression coverage for already-aligned device tensors
• verifies that cuvsDatasetMakePaddedView() remains a non-owning alias

Testing

Added/updated C++ and Python regression tests covering the aligned-device case and padded-view behavior.

Local static, Cython, formatting, and Python checks passed. CUDA/C++ runtime tests could not be run locally because the development environment does not have a CUDA toolkit or GPU, so those checks are left to project CI.

Fixes #2482

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[Sharon05564]

Hi! The label checker is currently failing because the PR is missing the required category and breaking-change labels. Could a maintainer please add the appropriate labels? This is a bug fix and is intended to be non-breaking. thanks!