Add table design doc

[jacquesqiao]

fix: #9183
related issue: #9068

[typhoonzero]

Sorry, I'm afraid we don't need this type of variable for the following reason.

1. One tensor distributed on multiple servers could do the same thing. Because the size of the table can be determined before training, distribute this tensor by dim0 and lookup the id by calculating the index boundaries of the servers.
2. If it's for an outside server, we'll only need to merge the keys to a SelectedRows variable.

[jacquesqiao]

The main problem is that in large scale embedding, the feature_id(int64_t) is discrete in a very big range, it's hard to represent it by Tensor or SelectedRows, for both of them use a dense tensor that the shape of memory should be decided before running.

So we need a new Variable that works like a key-value store or a map. It can:

1. get value by key.
2. store data that key is not continuous.

[typhoonzero]

@jacquesqiao I see, I figure it out now, thank you very much I'll look into the design.

[typhoonzero]

I've talked to @jacquesqiao today that we can use SelectedRows as the sparse weight table directly and decide which server should an id be put on by id % server_count.

[jacquesqiao]

yes, SelectedRows can support the need for the distributed lookup table. It's structure is

struct SelectedRows {
  Vector<int64_t> rows_;
  std::unique_ptr<Tensor> value_{nullptr};
  int64_t height_;
}

rows_ can be used to store the ids with the same shard_id
value_ store a dense matrix, user can get the value of a id using id's offset in the rows_.