Delete copy of quantized SDPA in torchao/experimental (pytorch#13924)

Summary:
X-link: pytorch/ao#2952


This deletes the SDPA/matmul kernels in torchao/experimental.

These kernels were already migrated to torchao/csrc/cpu/torch_free_kernels in the first diff, but we needed to keep them around in torchao/experimental to do a pin bump in ExecuTorch OSS after the first diff lands.


Test Plan:
Imported from GitHub, without a `Test Plan:` line.

Rollback Plan:

Reviewed By: kimishpatel

Differential Revision: D81640227

Pulled By: metascroy

Author: metascroy

extension/llm/custom_ops/op_sdpa_impl.h

@@ -24,7 +24,7 @@
 #endif
 #include <executorch/extension/kernel_util/make_boxed_from_unboxed_functor.h>
 
-#include <torchao/experimental/kernels/cpu/interface/quantized_matmul.h>
+#include <torchao/csrc/cpu/torch_free_kernels/interface/quantized_matmul.h>
 
 namespace torch {
 namespace executor {

extension/llm/custom_ops/targets.bzl

@@ -14,7 +14,7 @@ def _get_quantized_sdpa_deps():
     if runtime.is_oss:
         return []
     else:
-        return ["//pytorch/ao/torchao/experimental/kernels/cpu/interface:interface"]
+        return ["//pytorch/ao/torchao/csrc/cpu/torch_free_kernels/interface:interface"]
 
 def _get_quantized_preproc_flags():
     if runtime.is_oss: