Adding small python wrapper for multi_tensor_apply

apex/multi_tensor_apply/__init__.py

+from .multi_tensor_apply import MultiTensorApply
+
+multi_tensor_applier = MultiTensorApply(2048*32)
+

apex/multi_tensor_apply/multi_tensor_apply.py

+import torch
+
+class MultiTensorApply(object):
+    available = False
+    warned = False
+
+    def __init__(self, chunk_size):
+        try:
+            import amp_C
+            MultiTensorApply.available = True
+            self.chunk_size = chunk_size
+        except ImportError as err:
+            MultiTensorApply.availble = False
+            MultiTensorApply.import_err = err
+
+    def check_avail(self):
+        if MultiTensorApply.available == False:
+            raise RuntimeError(
+                "Attempted to call MultiTensorApply method, but MultiTensorApply "
+                "is not available, possibly because Apex was installed without "
+                "--cpp_ext --cuda_ext.  Original import error message:",
+                MultiTensorApply.import_err)
+
+    def __call__(self, op, noop_flag_buffer, tensor_lists, *args):
+        self.check_avail()
+
+        return op(self.chunk_size,
+                  noop_flag_buffer,
+                  tensor_lists,
+                  *args)