[Bugfix] Replace thread binding detector in LayoutInference Pass

[LeiWang1999]

CPU support introduced a flag skip thread binding into LayoutInference Pass, while it depends on buffer allocations to detect whether it's a cuda like device, which is not efficient and may introduce some bugs.

This pull request includes significant changes to the format.sh and src/transform/layout_inference.cc files, focusing on removing clang-tidy support and refactoring the layout inference logic. Below are the most important changes:

Removal of clang-tidy support:

• format.sh: Removed the entire section related to clang-tidy checks, including the installation check, function definitions, and script logic for running clang-tidy on files.

Refactoring layout inference logic:

• src/transform/layout_inference.cc: Replaced the AllocateCollector class with ThreadBindingCollector to focus on collecting thread bindings instead of memory allocations. Removed functions related to shared memory and local fragment checks.
• src/transform/layout_inference.cc: Updated the LayoutInference function to use ThreadBindingCollector and simplified the logic to determine if thread partitioning should be skipped based on the presence of thread bindings.