Update llama cmake for custom ops (pytorch#10176)

Fix pytorch#10166

target_link_options_shared_lib has different behavior for macos and
other platforms; seems like whole-archive was causing issues.

Test Plan
`sh .ci/scripts/test_llama_torchao_lowbit.sh`

https://github.com/pytorch/executorch/blob/main/examples/models/llama/README.md#running-with-low-bit-kernels

Author: lucylq

examples/models/llama/CMakeLists.txt

@@ -111,7 +111,8 @@ target_link_options_shared_lib(quantized_ops_lib)
 list(APPEND link_libraries quantized_kernels quantized_ops_lib)
 
 if(EXECUTORCH_BUILD_KERNELS_CUSTOM)
-  list(APPEND link_libraries $<LINK_LIBRARY:WHOLE_ARCHIVE,custom_ops>)
+  target_link_options_shared_lib(custom_ops)
+  list(APPEND link_libraries custom_ops)
 endif()
 
 if(EXECUTORCH_BUILD_TORCHAO)

examples/models/llama/README.md

@@ -431,7 +431,7 @@ cmake -DPYTHON_EXECUTABLE=python \
     -DEXECUTORCH_BUILD_EXTENSION_DATA_LOADER=ON \
     -DEXECUTORCH_BUILD_EXTENSION_MODULE=ON \
     -DEXECUTORCH_BUILD_EXTENSION_TENSOR=ON \
-    -DEXECUTORCH_BUILD_XNNPACK=ON \
+    -DEXECUTORCH_BUILD_XNNPACK=OFF \
     -DEXECUTORCH_BUILD_KERNELS_QUANTIZED=ON \
     -DEXECUTORCH_BUILD_KERNELS_OPTIMIZED=ON \
     -DEXECUTORCH_BUILD_KERNELS_CUSTOM=ON \