DG2 supports for XeTLA SDP example

[DDEle]

Type of Change: feature

Description

DG2 supports for XeTLA SDP example and some more improvements:

• In addition to the sdp example, also enables example 01 && 02 with minimal changes in the example code
  • by templating the default tunning options
• Generalized slm size checking for different gpu_arch
• Added some alias for dict checking (e.g. data_type_a_t<dict> ==> typename dict::template find_elem_t<tune_key::data_type_a>::type) (include/kernel/default_config/common.hpp)
• Added a tunning option key param_optimizer_mode which could have value of param_optimizer_mode::keep_shape (default to param_optimizer_mode::full) to prevent decision_tree_optimizer changing tile sizes (include/kernel/default_config/common.hpp)
• Replaced the implementation of some selector_t from deriving to type aliasing so it won't be a barrier when debuging with XETLA_PRINT<xxx_selector_t<...>>()
• Added an arch dispatcher running code with the current gpu_arch on current machine (probably not compatible with AOT)
• Fixed some typos
• Added editorconfig so that files won't be detected as 8-space-indentation

How has this PR been tested?

Compilation on linux-2024.0, linux-2024.1 and win-2024.0

cmake -GNinja .. && ninja

Examples on Arc and PVC (with no error):

cmake -GNinja .. && ninja gemm_universal && ./examples/01_gemm_universal/gemm_universal
cmake -GNinja .. && ninja basic_gemm && ./examples/02_basic_gemm/basic_gemm
cmake -GNinja .. && ninja sdp && ./examples/08_scaled_dot_product_attention/sdp

Unit tests on PVC:

cmake -GNinja .. && ninja && ctest -V

Both on this branch and the base branch failed on the following cases:

• 8 - gemm_int4_dequantization (Failed)

  but it is ok on 0x56a0

• 9 - gemm_dg2_int4_dequantization_bias (Failed)

  but it is ok on 0x56a0

• 26 - tile_load_store (Subprocess aborted)

Dependency Change?

No

[DDEle]

Just force-pushed to avoid reorganizing the kernel function of example 08.