Related issue = #5511

The PR seems a bit messy but I can't split it to smaller PRs because these changes must be done together to make the program run.

This PR roughly does things as follows:

1. Adds link_context_data and function link_compile_data in TaichiLLVMContext for linking.
2. Replace struct_module with struct_modules in ThreadLocalData. The struct_modules is a map that stores struct modules of every SNodeTree (The ID of the SNodeTree is the key of the map).
3. Let get_runtime_function and get_struct_function return a function declaration rather than the function definition.
4. Create a new module when starting to construct a struct module or a kernel module instead of cloning the struct module.
5. Support AOT and WASM.
6. Fix tests.

compile_kernel_to_module returns a vector with only one element now. The kernel cache contains only one module now, so it doesn't need to be stored in a vector anymore. I will refactor them later.