[BugFix][LLVM] Fix insertDeclare API mismatch for ROCm-bundled LLVM 20

[swjng]

Problem

Closes #18709.

ROCm ships its own LLVM 20 build that still uses the pre-LLVM-20 Instruction* overload of DIBuilder::insertDeclare, while upstream LLVM 20 switched to BasicBlock::iterator. The previous guard #if TVM_LLVM_VERSION >= 200 in codegen_llvm.cc broke ROCm users because their LLVM reports version 20 but lacks the new overload, causing a compile error like:

```
error: no matching function for call to
'llvm::DIBuilder::insertDeclare(..., llvm::BasicBlock::iterator)'
note: candidate: insertDeclare(..., llvm::Instruction*)
```

Fix

Replace the version-number guard with a CMake check_cxx_source_compiles feature probe in cmake/utils/FindLLVM.cmake. The probe compiles a small test program that calls insertDeclare with BasicBlock::iterator. The result is exposed as the preprocessor macro TVM_LLVM_INSERTDECLARE_USES_ITERATOR.

cmake_push_check_state/cmake_pop_check_state is used to avoid leaking CMAKE_REQUIRED_* variables into the rest of the build.

Three call sites in src/target/llvm/codegen_llvm.cc are updated to use #if TVM_LLVM_INSERTDECLARE_USES_ITERATOR instead of #if TVM_LLVM_VERSION >= 200.

Behavior by configuration:

LLVM build	Probe result	Macro set	Code path
Upstream LLVM >= 20 with new overload	Compiles	Yes	BasicBlock::iterator
ROCm LLVM 20 (old overload)	Fails	No	Instruction*
Any LLVM < 20	Probe skipped	No	Instruction*

Testing

Tested locally with LLVM 20.1.8 (Ubuntu package). The probe correctly detects Failed (uses Instruction* path), libtvm.so builds cleanly, and all 39 LLVM codegen tests pass:

```
tests/python/codegen/test_target_codegen_llvm.py 39 passed
```

[gemini-code-assist]

Code Review

This pull request introduces several fixes and improvements to the TVM build and compilation pipelines. It adds CMake feature detection for LLVM 20's DIBuilder::insertDeclare to handle differences between upstream and ROCm-bundled versions. In the Relax VM build, it now prefers target-specific pipelines when using the "default" configuration. The compile_relax integration for MetaSchedule was refactored to ensure a correct pass ordering, including DLight fallback scheduling. Additionally, safety checks were added to TIRX passes to handle functions without defined targets. Feedback suggests improving the target-specific pass detection in the MetaSchedule integration to support backends beyond CUDA, avoiding manual duplication of pass lists, and ensuring CMake variable state is preserved during feature detection.