Acquire the GIL in lazy init_extension

Bindings declared with call_guard<gil_scoped_release> (e.g. fp8_transpose)
can be the first transformer_engine_torch call in a process. init_extension()
then runs py::module_::import without holding the GIL and segfaults inside
PyImport_ImportModule. Reacquire the GIL inside the one-time initializer.

Signed-off-by: zhihaow6 <zhihaow6@illinois.edu>

Author: xiuhu17

transformer_engine/pytorch/csrc/extensions/pybind.cpp

@@ -121,6 +121,7 @@ void init_grouped_tensor_extension() {
 
 void init_extension() {
   std::call_once(extension_init_flag, []() {
+    pybind11::gil_scoped_acquire gil;
     init_float8_extension();
     init_mxfp8_extension();
     init_float8blockwise_extension();