[KDA][AMD]for kda kernel,fix core dump on AMD GPU and tune the config for AMD branch

[binding7012]

Original code run on AMD gpu would core dump. this pr fix it, and also tune the config for AMD gpu

[gemini-code-assist]

Code Review

This pull request introduces AMD GPU support by adjusting autotune configurations and implementing Python-level state transposition to meet hardware constraints. It also includes a bug fix in the fused recurrent KDA kernel to properly mask memory loads. The review feedback suggests refactoring duplicated kernel launch logic to improve maintainability, avoiding performance issues caused by sequence length specialization on AMD, and maintaining consistency in autotune keys across platforms. Additionally, the reviewer recommends clarifying the impact of disabling transposed layouts for KDA on ROCm.

[binding7012]

Hi @yzhangcs pls help to review this PR, thanks~

[zhiyuan1i]

Thanks for the PR! ROCm support is definitely on our radar. We're currently working on a more unified backend abstraction design to avoid scattering hardware-specific branches across kernels. We'd prefer to land KDA's AMD support through that framework rather than merging this directly. Let's keep this open as a reference for now?