Feature/nmr mlp

[tomoqt]

added a simple 2x MLP to encode raw NMR instead of interpolating them to 1k. both NMRs are projected to memory_dim and concatenated over seq_len.
Transformer decoder should be ok, as uses whatever seq_len comes in.
note this doesn't cntain info about the multiplets (for now), plus the projection is pretty sharp, especially for low embed_dim (maybe should be more bottleneck than expanding dimension MLP...). note that this way we use the full embed_dim for IR, hence modalities are less separated between different attention heads.
setting use_mlp_for_nmr=False gives back regular functionality.
maybe a next step could be validating the effect of regular vs this. eventually we can:
-keep convnext, change the backbone HPs to downsample more and get around regular seq_len as the IR, then pad and concatenate as usual over embed_dim.

• change preprocessing over NMR (FT? CWT?discretize space of peaks like the paper did?)