Python: grid_spacing & time_offset double

[ax3l]

In pybind11, overloads on types are order-dependent (first wins).
pybind/pybind11#1512

We specialize double here generically and cast in read if needed (see #345 #1137).

Later on, we could add support for 1D numpy arrays with distinct type.

[franzpoeschel]

Is there a reason not to specify long double generically instead of double? With this PR, don't you get a narrowing conversion if the Series stores these attributes as a long double type?

[ax3l]

So in Python there is only a dynamic-bitness builtin float type. We could use long double here to case up to the platform-specific largest type, but I think I was conservative in my design here because long double differs from platform to platform.

I think the most general thing that we could do for the getter:

• write a lambda switch loop that checks at runtime the type of the attribute
• return a std::variant<float, double, long double>
  and for the setter:
• overload for generic Python double type and
• numpy scalar

For now, this feels like a bit too detailed since there is most likely a conversation when going to Python builtin float anyways.

[franzpoeschel]

With the current status of this PR, I think the following situation might happen:

openPMD dataset        ---> openPMD Python API               ---> Python script
(contains long double)      (uses Mesh::gridSpacing<double>)      (platform supports long double, but value was cast to double)

Your point of cross-platform consistency is valid, so I think we can go forward with this PR. We should just maybe be aware that this leaves stored long double values accessible only via C++.

[ax3l]

Yes, that's the problem we need to generalize 👍