Revert "Revert "Revert of #3822 "Revert "implementing sym_numel_custom"" (#3826)" (#3835)" (#3837)

This reverts commit 3af0873.

Author: Krovatkin

torch_xla/csrc/tensor_impl.cpp

@@ -120,6 +120,15 @@ c10::SymIntArrayRef XLATensorImpl::sym_sizes_custom() const {
                              sizes.size());
 }
 
+c10::SymInt XLATensorImpl::sym_numel_custom() const {
+  auto sym_sizes = sym_sizes_custom();
+  c10::SymInt prod{1};
+  for (auto s : sym_sizes) {
+    prod *= s;
+  }
+  return prod;
+}
+
 c10::SymIntArrayRef XLATensorImpl::sym_sizes() const {
   // it isn't strictly necessary to delegate to `sym_sizes_custom`
   // however, it's consistent with pytorch core

torch_xla/csrc/tensor_impl.h

@@ -35,6 +35,7 @@ class XLATensorImpl : public c10::TensorImpl {
   at::IntArrayRef sizes_custom() const override;
   c10::SymIntArrayRef sym_sizes() const override;
   c10::SymIntArrayRef sym_sizes_custom() const override;
+  c10::SymInt sym_numel_custom() const override;
   at::IntArrayRef strides_custom() const override;
 
   int64_t dim_custom() const override;