apply index_type on value

[piever]

Fix #203 by running the index_type computation by recursing tuples rather than tuple types.

@oschulz I think the basic mistake was to compute index_type on types rather than values. Does this PR fix your type instability problem?

@ranocha this PR should not regress the performance improvements you introduced in #194, but I guess it'd be good to double check it in practice if you have a simple way to run the performance benchmarks on Trixi based on this PR

[ranocha]

This PR doesn't seem to impact the performance of the benchmark setups I looked at in #194.

[oschulz]

@piever thanks, I'll try the PR!

I'm currently using this workaround in BAT.jl:

_structvector_indexstyle(contents::Tuple) = _structvector_indexstyle_impl(map(IndexStyle, map(typeof, contents)))
_structvector_indexstyle_impl(::NTuple{N,IndexLinear}) where N = Int
_structvector_indexstyle_impl(::NTuple{N,IndexStyle}) where N = CartesianIndex{1}

It also uses the values - it's a bit too simple (limited to vectors), but maybe the basic map-IndexStyle-over-contents approach could get rid of some of the recursion in the updated StructArrays code in this PR?

[oschulz]

@piever sorry for the delay - did a few additional tests - looks like this PR fixed the problem completely. Thanks!

[piever]

Great, thanks for double checking!

[JoaoAparicio]

Now index_type some times takes values:

StructArrays.jl/src/structarray.jl

Line 26 in 99f0556

index_type(components::NamedTuple) = index_type(values(components))

and some times takes types:

StructArrays.jl/src/lazy.jl

Line 95 in 99f0556

index_type(::Type{LazyRows{T, N, C, I}}) where {T, N, C, I} = I

Seems strange.

[piever]

It's true that it is a bit confusing. It turns out that we can just use Base.IndexStyle directly and it seems to infer fine, I've opened #288 to address that.