Inconsistency in naming of extracted representations

Hello,

Really nice work! I think I encountered a minor inconsistency in the naming of the extracted representations. Specifically, after running the guideline in `pretrain_repr/openfold`, the extracted representations have the format:

`{name}.edge_repr.recycle3.npy` and `{name}.node_repr.recycle3.npy`

So, when running `src/eval.py` the code has a bug in `src/data/full_atom/feat_loader.py` :

```
if self.node_size > 0:
node_repr_path = (
self.data_root
/ f"{prefix}"
/ f"{prefix}_recycle{self.num_recycles:d}_single_repr.npy"
```
because it searches for a representation with a different naming. To make it work I had to do the following:

Replace in line 53:

```
node_repr_path = (
self.data_root
/ f"{prefix}"
/ f"{prefix}_recycle{self.num_recycles:d}_single_repr.npy"
)
```
with

```
node_repr_path = Path(
str(self.data_root / f"{prefix}")
+ f".node_repr.recycle{self.num_recycles:d}.npy"
)
```

Replace in line 67:
```
edge_repr_path = (
self.data_root
/ f"{prefix}"
/ f"{prefix}_recycle{self.num_recycles:d}_pair_repr.npy"
)
```
with
```
edge_repr_path = Path(
str(self.data_root / f"{prefix}")
+ f".edge_repr.recycle{self.num_recycles:d}.npy"
)
```

After these changes, it worked on my case.