UDF EdgeBatch.edges() confuses original graph and compute graph when message passing

## 🐛 Bug

Edge-wise User-defined Function [EdgeBatch.edges](https://docs.dgl.ai/generated/dgl.udf.EdgeBatch.edges.html#dgl.udf.EdgeBatch.edges)() raise error as it confuses original graph and compute graph.

It happens when we use UDF as the message function inside of `pull()`, `push()` and `send_and_recv()` as they all call `DGLHeteroGraph._create_compute_graph()`. It extracts a subgraph named compute graph from the original one in order to compute message passing.

However, in this function, it sets the edges' id `_eid` according to the original graph instead of the relabelled ones, making edge index `_eid` exceeds the total number of edges in the compute graph. Thus, when we call `EdgeBatch.edges()` in the UDF, trying to find the edges with `_eid` in the compute graph instead of the original one causes `Invalid edge ID` error.

## To Reproduce

Run the code below. `pull()`, `push()` and `send_and_recv()` all raise this error.

```python
import dgl, torch, dgl.function as fn
g=dgl.graph(([0,1,2,3],[3,2,1,4]))

def msg_fn(edges):
src, dst, _ = edges.edges()
return {'m': torch.zeros(len(src), 1)}

g.pull([1], msg_fn, fn.sum('m','h'))
g.push([1], msg_fn, fn.sum('m','h'))
g.send_and_recv([1], msg_fn, fn.sum('m','h'))
```

## Expected behavior

Error `dgl._ffi.base.DGLError: Invalid edge ID 1` is raised.

## Environment

- DGL Version: 0.9.0
- Backend Library & Version: 1.12.1
- OS: Mac OS 13.0
- How you installed DGL: conda 4.13.0
- Python version: 3.10.4