[Feature request] Support reservoir sampling or GPU compact_blocks to accelerate TGN

## 🚀 Feature

Support for using gpu to train tgn model

## Motivation

There is an example of tgn here: https://github.com/dmlc/dgl/tree/master/examples/pytorch/tgn. But it can only use cpu to train the model so it's much slower than other implementation(more than 10x times).

## Alternatives

## Pitch

It seems to be related about TemporalEdgeDataLoader class which leads it difficult to implement a GPU version. It will be very nice if dgl team can provide a GPU version of tgn.

## Additional context

I have tried to assign graph to GPU. I think the possible problem is this: dgl dataloader uses built-in function compact_graphs() in transform.py which needs all graphs on CPU.