fix/rocm properties export

[GitGlimpse895]

ROCMProperties was fully implemented in device.py with its own docstring and example, but was never wired up end-to-end:

• Device.properties was typed as CUDAProperties | None, silently rejecting ROCMProperties at the type-system level.
• Device.__post_init__ validated CUDAProperties placement but had no symmetric guard for ROCMProperties, allowing invalid combinations like Device(type="cuda", properties=ROCMProperties(...)) to pass silently.
• ROCMProperties was never exported from kernels/layer/__init__.py or the top-level kernels/__init__.py, making from kernels import ROCMProperties fail despite being shown in the class docstring example.

This PR fixes all three gaps across three files, completing the ROCm capability-range kernel selection feature that was already wired up in _ROCMRepos.

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[GitGlimpse895]

Hey @sayakpaul — wanted to get your eyes on this if you have a moment! This PR wires up ROCMProperties end-to-end — fixes the type annotation on Device.properties, adds a symmetric validation guard in __post_init__, and exports ROCMProperties from both kernels/layer/__init__.py and the top-level kernels/__init__.py. It's a pretty contained fix, but curious what you think — does this approach look right to you, or would you handle the validation guard differently?

[danieldk]

Looks good, thanks! Sorry for the delay.