Fix NVIDIA#633: Bump the LLVM version to 16.0.6.

This changes the internal llvm (tpls/llvm) to point to a SHA on the release/16.x
branch. This SHA corresponds to version 16.0.6.

Documentation should be updated to refer to LLVM 16.0.6 as the latest version
supported.

Add workaround for unsigned integer constants as template arguments.

Move the subdirectory up in the file to avoid it getting "lost". It's not clear
what the bug is towards the end of the file that was causing it to be omitted
in my workspace setup.

lib/Frontend/nvqpp/ConvertDecl.cpp

@@ -157,7 +157,12 @@ bool QuakeBridgeVisitor::interceptRecordDecl(clang::RecordDecl *x) {
         auto templArg = tempSpec->getTemplateArgs()[0];
         assert(templArg.getKind() ==
                clang::TemplateArgument::ArgKind::Integral);
-        std::int64_t size = templArg.getAsIntegral().getExtValue();
+        auto getExtValueHelper = [&](auto v) -> std::int64_t {
+          if (v.isUnsigned())
+            return static_cast<std::int64_t>(v.getZExtValue());
+          return v.getSExtValue();
+        };
+        std::int64_t size = getExtValueHelper(templArg.getAsIntegral());
         if (size != static_cast<std::int64_t>(std::dynamic_extent))
           return pushType(quake::VeqType::get(ctx, size));
       }

unittests/CMakeLists.txt

@@ -133,6 +133,8 @@ target_link_libraries(test_spin
   gtest_main)
 gtest_discover_tests(test_spin)
 
+add_subdirectory(plugin)
+
 # build the test qudit execution manager
 add_subdirectory(qudit)
 add_executable(test_qudit main.cpp qudit/SimpleQuditTester.cpp)
@@ -183,5 +185,3 @@ if (CUDAQ_ENABLE_PYTHON)
     TEST_SUFFIX _Sampling PROPERTIES ENVIRONMENT "PYTHONPATH=${CMAKE_BINARY_DIR}/python")
 endif()
 
-add_subdirectory(plugin)
-