PR tensorflow#23426: [XLA:GPU] set CUBLASLT command buffer as default enabled

Imported from GitHub PR openxla/xla#23426

Copybara import of the project:

--
3d2f748916dc6a9c72c681f9bd9fc46f03d396ee by Shawn Wang <shawnw@nvidia.com>:

set CUBLASLT command buffer as default enabled

Merging this change closes tensorflow#23426

PiperOrigin-RevId: 734372857

Author: shawnwang18

third_party/xla/xla/debug_options_flags.cc

@@ -116,6 +116,7 @@ DebugOptions DefaultDebugOptionsIgnoringFlags() {
 
   opts.add_xla_gpu_enable_command_buffer(DebugOptions::FUSION);
   opts.add_xla_gpu_enable_command_buffer(DebugOptions::CUBLAS);
+  opts.add_xla_gpu_enable_command_buffer(DebugOptions::CUBLASLT);
   opts.add_xla_gpu_enable_command_buffer(DebugOptions::CUSTOM_CALL);
   opts.add_xla_gpu_enable_command_buffer(DebugOptions::CUDNN);
   opts.set_xla_gpu_graph_min_graph_size(5);