Add function to get element count from tensor.

[qingqing01]

Fix #3914

[JiayiFeng]

Why not use product() ?

[qingqing01]

@Canpio framework::product(tensor->dims()) is used in many places. A common interface is convenient.

[QiJune]

It seems that the size of tensor is often used, so make a cache is necessary.

[QiJune]

LGTM

[JiayiFeng]

LGTM.
As a cache of element number, numel had appeared in a very early design of Tensor. However, then it was removed because we were not sure whether it is necessary. During the recent developing, many colleagues have voiced that writing framework::product(tensor.dims()) is tedious and inefficient, for tensor's dims never change in the majority of cases. So, I think adding numel is reasonable. @wangkuiyi

[pkuyym]

Great, I think this PR is very helpful. When converting a tensor to a matrix, we often calculate the column number using framework::product(ddim). If we have an cache variable, things will be simplified.

[wangkuiyi]

I agree that we add Tensor::numel() as it is such a commonly used feature.

I noticed that in the tensor implementation, we use Tensor::numel_, but in other places, we call Tensor::numel(). This implies that that is a cache -- the former, and the latter returns the former. It would be great if we don't expose such implementation details to code readers. It looks to me that we can just call Tensor::numel() everywhere.

[qingqing01]

Done.

[wangkuiyi]

/*! A cache of the number of elements in a tensor. Would be 0 for an uninitialized tensor. */

[qingqing01]

Done.

[QiJune]

LGTM