FNO and TFNO did not the FNO/TFNO blocks

[AnnihilatorChess]

Hello,

After some weird behavior of the FNO i discovered that the FNO blocks are not used in both models. The blocks are called but the output is not assigned to x. The same is true for the tfno.

I tested this wrong implementation of the FNO and managed to reproduce the benchmarks for the turbulent_radiative_layer_2D dataset of the paper (one-step VRMSE: ~0.5). After applying this change to actually use the FNO blocks, I tested the FNO with the default settings of the_well for 100 epochs (adjusted LR to 1e-3 and batch_size=8 to fit on my GPU) and achieved VRMSE of ~0.3. This makes me think that the benchmark results in the paper for the turbulent_radiative_layer_2D dataset were also computed with this incorrect implementation.
I can share the results of both runs (wrong implementation and correct implementation) if you want.

On another note that has nothing to do with this pull request. The test set performance is computed with the most recent and not the best model (on the val set). I am not sure whether this is a design choice or if this was overlooked.

[mikemccabe210]

Thanks for catching that, we'll include these modifications as we rerun some benchmarks to make sure whatever checkpoints are public match posted numbers. (@kazewong)

Setting off the CI for now, but we can merge this as soon as ruff is happy.

[mikemccabe210]

Great, looks like everything passed off the bat! Thanks again for catching this @AnnihilatorChess .

[mikemccabe210]

Also, answer to the second point - I would classify that as part of the bug-to-feature pipeline in that we're aware it is the case, but it was not originally intentional and was largely an artifact of co-building the validation and resume pipelines.

[AnnihilatorChess]

Is there any update on the benchmarks for the corrected versions?