[Refactor] Enhance MHA kernel with support for different architectures and data types

[Rachmanino]

This pull request introduces several improvements to the test configuration and code organization for multi-head attention (MHA) kernels, as well as a new utility function for device architecture detection. The main changes include adding support for configurable data types in the MHA tests, refactoring imports for better modularity, and introducing an is_hopper utility function.

Test configuration improvements

• Added a --dtype argument to the MHA test script (tests/test_mha.py), allowing selection between float16 and bfloat16 data types, and updated the test function to use this argument. Also adjusted defaults for --causal and --tune arguments to improve usability. [1] [2]
• Changed the default value of the --tune argument in the MHA decode test script (tests/test_mha_decode.py) to False for consistency with other test scripts.

Code organization

• Refactored kernel imports in top/__init__.py and top/kernel/__init__.py to import MHADecodeKernel from its own module (mha_decode) instead of mha, improving modularity and maintainability. [1] [2]
• Added an __all__ declaration to top/kernel/mla.py to explicitly define the public API of the module.

New utility function

• Added a new is_hopper function in top/utils/utils.py to detect if the current CUDA device is of the Hopper architecture (compute capability 9.0).