Float32 high resolution spectral transform about speedyweather.jl

[milankl]

This seems to be an issue with the calculation of Legendre polynomials in Float32. At the moment, for recompute_legendre=false (default) we precompute the polynomials in Float64 and the convert them down. However, when using gridded the default was to recompute the polynomials as they wouldn't be stored anyway. AssociatedLegendrePolynomials.jl seems to have a bug with Float32 and high resolution (my guess), this is raised therein jmert/AssociatedLegendrePolynomials.jl#27

[milankl]

Legendre computation for orography now in Float64 with #105

[milankl]

#117 now has Float64 precomputation of the Legendre polynomials as default (and then converts to the actually used number format), so this issue is resolved as long as we don't run at higher than T3000 😉 . The recomputation is now always done in Float64, which triggers parts of the integration in higher precision but converts down. This is now the default for gridded and spectral.