Replace ReduceToAssign with Reduce

[reminisce]

This PR should fix this issue.
#6324 (comment)

Test script:

 import mxnet as mx
 import time
 
 data = mx.sym.Variable('data')
 sym = mx.sym.broadcast_to(data, shape=(62, 2048, 14, 14))
 exe = sym.simple_bind(ctx=mx.gpu(0), data=(62, 2048, 1, 1)) 
 
 exe.forward(is_train=True)
 exe.backward(out_grads=exe.outputs[0])
 time.sleep(3)  #https://github.com/dmlc/mxnet/issues/1137

@piiswrong @eric-haibin-lin @madjam please review.

[piiswrong]

Did you remove all reference to ReduceToAssign? If so could you remove the function alltogether?

[piiswrong]

?

[reminisce]

This line prevents me from using this macro in broadcast_reduce_op.h. So I had to remove it.

I will remove the function definitions of ReduceToAssign.