Bump NeuralOperators.jl to v0.6

[ChrisRackauckas]

No description provided.

[ChrisRackauckas]

@avik-pal I threw claude Opus on this:

Summary

  I've investigated the NeuralOperators.jl v0.6 update for NeuralPDE.jl.
  Here's what I found:

  Key Breaking Changes in NeuralOperators.jl v0.6:

  1. Removed permuted kwarg (default now corresponds to previous
  permuted=Val(true))
  2. Removed direct dependency on FFT packages
  3. Switched primary backend to Lux + Reactant

  NeuralPDE.jl Status:

  - Good news: NeuralPDE.jl doesn't use the permuted parameter, so that
  breaking change doesn't affect it
  - Good news: The Project.toml already specifies compatibility with
  NeuralOperators v0.6
  - Issue found: There's a bug in NeuralOperators.jl v0.6 where DeepONet
  fails with a dimension mismatch error when performing matrix
  multiplication between 2D and 3D arrays

  The Bug:

  DeepONet in v0.6 attempts to multiply:
  - Branch output (after transpose): 2D matrix of shape (5, 10)
  - Trunk output: 3D array of shape (10, 10, 1)

  This operation is not defined in Julia and causes a MethodError.

  Recommendation:

  The NeuralPDE.jl code itself doesn't need any changes for v0.6
  compatibility. However, there's a bug in NeuralOperators.jl v0.6 that
  prevents DeepONet from working correctly. This issue should be reported
  to the NeuralOperators.jl repository so it can be fixed in a patch
  release (v0.6.1).

Are you aware of this? Is it missing a breaking change?

[avik-pal]

I updated DeepONet to match the original paper defn as part of v0.6 release. The previous one was needed because we couldn't fuse ops, but with reactant as the default those manual tunings aren't needed.

If you point me to which example in NeuralPDE is failing, I can update that. Also in general, we should probably hold off updating NeuralPDE to latest NO till this is ready for Reactant support. v0.5 performance for non-Reactant backends should be better