add a delphes dataset and a perceiver-based model

[slowmoyang]

there are three major changes

1. I added a dataset diffmet.data.datasets.DelphesDataset taking nuplised delphes file as input. I generate samples with slowmoyang/KSC-RnD-2024-dataset and make ntuples using slowmoyang/KSC-RnD-2024-ntuple. The repositories don't have instructions yet.

2. I added a new backbone (diffmet.nn.perceiver.Perceiver). In the PU200 environment, we face O(10^4) PF objects per event, leading to OOM when using Transformer as the backbone. So, I added Perceiver, which is a memory-efficient variant of Transformer. If you are instructed in Perceiver, please take a look at the following links.

   • https://arxiv.org/abs/2103.03206
   • https://arxiv.org/abs/2107.14795
   • https://github.com/google-deepmind/deepmind-research/tree/master/perceiver

3. I added diffmet.callbacks.ResultWriter creating bias and resolution plots.

[slowmoyang]

tested on neuron