Remove dependency on torch_scatter and torch_sparse

[RaulPPelaez]

Removes dependency on torch_sparse by implementing message passing operations with manually written ones.

Removes dependency on torch_scatter by using torch.index_add and torch.scatter_reduce.

[RaulPPelaez]

cc @guillemsimeon @AntonioMirarchi. Correctness tests are passing, but for this PR I had to reimplement the message passing of ET, T, and GN.

I do not want to merge it before a long train has been carried out to compare nothing broke.
Can you recommend a good test yaml for this? Please post it here and I will run it with/without this PR.

[guillemsimeon]

For the ET, the MD17 is fine. For the T, I would directly remove the architecture from TorchMD-NET, and I do not know any case where it has been used. For the GN, you can perhaps use the ones here https://arxiv.org/pdf/2212.07492.pdf

[RaulPPelaez]

@giadefa, do you agree with removing the Transformer architecture altogether? Its less dev burden for sure...

[guillemsimeon]

Raul, could you also train TensorNet on QM9 U0 with seed 1 as part of these tests? The yaml is in the repo.

[RaulPPelaez]

I would say ET is correctly implemented:

[RaulPPelaez]

TensorNet also looks ok:

[RaulPPelaez]

GN also seems fine:

Not sure what makes GN be so much more accurate in terms of reproducibility than the other ones, though.
I am using this yaml:

activation: tanh
batch_size: 8192
inference_batch_size: 8192
dataset: Custom
coord_files: "chignolin_ca_coords.npy"
embed_files: "chignolin_ca_embeddings.npy"
force_files: "chignolin_ca_deltaforces.npy"
cutoff_upper: 12.0
cutoff_lower: 3.0
derivative: true
early_stopping_patience: 30
embedding_dimension: 128
lr: 0.0005
lr_factor: 0.8
lr_min: 1.0e-06
lr_patience: 10
lr_warmup_steps: 0
model: graph-network
neighbor_embedding: false
ngpus: -1
num_epochs: 100
num_layers: 4
num_nodes: 1
num_rbf: 18
num_workers: 8
rbf_type: expnorm
save_interval: 2
seed: 1
test_interval: 2
test_size: 0.1
trainable_rbf: true
val_size: 0.05
weight_decay: 0.0

[guillemsimeon]

it looks good to me