Configure LIBXSMM with default arguments (tensorflow#28)

* Fixed AVX-512 intrinsic implementation.

* OR'ed LIBXSMM_DNN_CONV_OPTION_OVERWRITE into convolution options, which folds zeroing the input buffer on first use. This removes the call to libxsmm_dnn_zero_buffer in case of LIBXSMM_DNN_COMPUTE_KIND_FWD.

* Rely on libxsmm_hash rather than std::hash. Brought xsmm_conv2d.cc up-to-date with TF/master.

* Code cleanup: use LIBXSMM_DNN_CONV_OPTION_WU_EXT_FILTER_REDUCE_OVERWRITE rather than assembling the option from separate flags.

* Avoid to destroy the handle in case of LIBXSMM_DNN_WARN_FALLBACK since the next iteration may double-delete the same handle. One would need to update the handle-cache to allow destruction at this place. However, all handles are destructed when TF terminates (cache cleanup).

* Rely on default configuration arguments, and thereby lower the dependence from LIBXSMM internals.

third_party/libxsmm.BUILD

@@ -11,19 +11,8 @@ exports_files(["LICENSE"])
 libxsmm_interface_arguments = "0 1"
 
 # Arguments to ./scripts/libxsmm_config.py, see that file for detailed description.
-#  ilp64: no
-#  big: no
-#  offload: no
-#  alignment [b]
-#  prefetch: 1 (auto)
-#  threshold: fallback to BLAS if n*m*k above this
-#  synchronize: yes
-#  jit: yes
-#  flags
-#  alpha = 1
-#  beta = 1
-#  gemm = 2
-libxsmm_config_arguments = "0 0 0 64 1 0 1 1 0 1 1 2"
+# rely on default arguments
+libxsmm_config_arguments = ""
 
 # Arguments to ./scripts/libxsmm_dispatch.py, see that file for detailed description.
 #  (dummy argument)