Add Lod Reset Operator

[kuke]

Resolve #4719

[qingqing01]

Now, the attributes are usually fixed during training, that is to say, it was passed to the network when creating the graph. I'm not sure whether the attributes can be changed during training in the future. And since the batch size and the sequence length may be varying during training, I think it's better to pass LoD as an input, not attributes. The input can be calculated by others operator.

[kuke]

Done

[qingqing01]

(LoDTensor)The -> (LoDTensor) The

[kuke]

Done

[qingqing01]

target_lod_in -> TargetLoD

[kuke]

Done

[kuke]

Thanks. Updated by following the comments

[kuke]

Done

[kuke]

Done

[qingqing01]

out->ShareDataWith(*in);

[kuke]

Done

[qingqing01]

d_x->ShareDataWith(*d_out);

[kuke]

Done

[qingqing01]

In fact, the LodTensor X is saved as with shape (6, 1) in our codes:

[1,
 2,
 3,
 4,
 5,
 6]

But [[1, 2], [3, 4, 5], [6]] is easy to understand, I wonder how to add this comments.

[kuke]

Modified

[qingqing01]

if (platform::is_gpu_place(ctx.GetPlace())) {
}

[kuke]

Done

[kuke]

Thanks

[kuke]

Modified

[kuke]

Done

[kuke]

Done

[kuke]

Done

[qingqing01]

LGTM.

[pkuyym]

LGTM