Enable megacore_dense by default (#5520) (#5531)

Summary:
This change enables megacore_dense by default to allow asynchorous cc
ops especailly for GSPMD.

Test Plan:
CI

Co-authored-by: Jiewen Tan <jwtan@google.com>

torch_xla/__init__.py

@@ -65,6 +65,7 @@ def _setup_default_env():
   _set_missing_env('TPU_ML_PLATFORM', 'PyTorch/XLA')
   if server_is_alive():
     _set_missing_env('XRT_START_LOCAL_SERVER', '0')
+  _set_missing_env('TPU_MEGACORE', 'megacore_dense')
 
 
 _fd, _tmp_fname = -1, ''
@@ -170,4 +171,4 @@ def _init_xla_lazy_backend():
 # TODO @wonjoo come up with a long term fix in Dynamo.
 torch._dynamo.config.automatic_dynamic_shapes = False
 
-from .stablehlo import save_as_stablehlo, save_torch_model_as_stablehlo
+from .stablehlo import save_as_stablehlo, save_torch_model_as_stablehlo