[Dev] Fix bugs for ROCm with our default TileLang Backend

[LeiWang1999]

This pull request includes significant changes to enhance the target detection functionality and streamline the codebase. The most important changes involve the introduction of a new auto_detect_target function, the removal of redundant code, and the consolidation of target detection logic.

Enhancements to target detection:

• bitblas/utils/target_detector.py: Introduced the auto_detect_target function to detect the computing target (CUDA or ROCm) based on the environment, replacing the previous auto_detect_nvidia_target function in various parts of the codebase. [1] [2] [3]

Codebase simplification:

• bitblas/__init__.py: Replaced auto_detect_nvidia_target with auto_detect_target in the import statement.
• bitblas/base/arch/__init__.py: Removed the temporary auto_infer_current_arch function and added the new auto_detect_target function. [1] [2] [3]
• bitblas/benchmark/operator/__init__.py: Updated the BitblasOperatorBenchmarkBase class to use auto_detect_target instead of auto_detect_nvidia_target. [1] [2]
• bitblas/cache/operator.py: Updated the load_global_ops_cache function to use auto_detect_target instead of auto_detect_nvidia_target. [1] [2]

Note that currently only support consistent precision, dequantize op is coming soon.

[LeiWang1999]

• Dequantize Kernel without fast decoding
• Dequantize Kernel with fast decoding

[LeiWang1999]

TileLang fine grained implementation should be renamed into mma, as we will introduce mfma template kernels.