Parallel processing doesn't work when generating training datasets using get_3d_lmdb.py

Hi authors,

Thanks for the great work. I ran into some issues when following the readme of Uni-Mol+ to generate the training dataset. And I hope to get some advice on that.

When I run `python ../get_3d_lmdb.py train`, it needs to take ~250 hrs (according to tqdm) to finish the dataset generation (number of CPU cores on our machine is 112). Then I used 10 molecules to test the speed and I found the speed of sequential processing and parallel processing is about the same. Then I narrow it down to the function [rdkit_3d_gen](https://github.com/dptech-corp/Uni-Mol/blob/def0b42b972eb063531bf7460ad063d0c0136ce4/unimol_plus/scripts/get_3d_lmdb.py#L133) (shown below), which blocks the speed up in multiprocessing.

```python
def rdkit_3d_gen(smile, seed):
mol = read_smiles(smile)
AllChem.EmbedMolecule(mol, randomSeed=seed, maxAttempts=1000)
mol = rdkit_mmff(mol)
pos = mol.GetConformer().GetPositions()
return mol
```

If I comment out `AllChem.EmbedMolecule(mol, randomSeed=seed, maxAttempts=1000)`, the speed of parallel processing can become normal.

I appreciate it if any suggestions on how to fix this issue. Looking forward to hearing from you.

Thank you