[LLDB][AMDGPU] Refine GPU core comparison framework

[thechenli]

Stacked on llvm#104.

• Move auto-load behavior into load_core(..., auto_load_debuginfo=False) instead of reading GPU_COMPARISON_AUTOLOAD from the test.
• Compare ROCLLDB CPU + GPU modules as one combined module list against ROCgdb's flat objfile list.
• Normalize debugger-specific module keys for .gnu_debugdata, memory code objects, file://...#offset=..., and [vdso].
• Keep local variables/register/thread count behavior unchanged outside the load-core plumbing.

Validation:

• python3 -m py_compile on the modified comparison framework files.
• Ran the comparison eval on /home/chenlii/mast_coredumps.jaexkvnih5h2popy.merged.
• Registers and thread count passed.
• Module comparison now reduces to the known ld.so outlier.
• Local variables passed because both debuggers reported zero locals for the selected GPU frame.

Results from roclldb_eval_jaex_once/comparison_dotest.log:25436:

• backtrace: failed only on frame 1 PC difference.
  GDB 0x7eb724868efc vs LLDB 0x7eb72486889c, same function/source line.

• modules: failed only on the known outlier:
  GDB modules: 1084, LLDB modules: 1070, Module 'ld.so' only in GDB.

• local_variables: passed.

• registers: passed, GDB registers: 120, LLDB registers: 120.

• thread_count: passed, GDB total: 2106, LLDB CPU: 1186, LLDB GPU: 920, LLDB total: 2106.

• placeholder: passed.