[Doc] Remove unnecessary layout annotation

[LeiWang1999]

This pull request includes several changes to improve the functionality and performance of the flash_attention and flash_decoding modules, as well as updates to the testing framework. The most important changes include modifications to the example_mha.py and example_mha_inference.py files, updates to the test_tilelang_tilelibrary_gemm.py file, and adjustments to the fragment.py file.

Flash Attention and Flash Decoding Updates:

• examples/flash_attention/example_mha.py: Commented out the T.annotate_layout line and adjusted parameters for flashattn to improve performance. [1] [2]
• examples/flash_decoding/example_mha_inference.py: Removed the T.annotate_layout line to streamline the code.

Testing Framework Enhancements:

• testing/python/tilelibrary/test_tilelang_tilelibrary_gemm.py: Added a comprehensive set of tests for GEMM operations across various data types and configurations, including float16, bfloat16, float32, float64, and int8.
• testing/python/kernel/test_tilelang_kernel_gemm.py: Re-enabled the tilelang.testing.main() call to ensure all tests are executed.

Codebase Simplification:

• tilelang/layout/fragment.py: Simplified the conditional check for forward_index to improve readability and maintainability.

Documentation:

• examples/flash_decoding/README.md: Added a new section titled "Flash Decoding" to provide better documentation and guidance.