Improve efficiency of summary metadata aggregation

[asnyv]

Currently the summary metadata is handled by a separate method in fmu-ensemble, which is about as heavy as the get_smry() itself, as it loads all the summary data.

In ecl2df it is now proposed to add this metadata directly to the pandas DataFrame.attrs for the dataset. The DataFrame.attrs is said to be experimental, so a bit risky to base on it, but there are some good opportunities. An alternative is something like returning a tuple of the df and a metadata dict instead of just the df if you set a flag.
At least, fmu-ensemble is planning to base this part of the code directly on ecl2df, so then we could get the same feature there.

On the webviz-side, the pandas.to_parquet() for portables doesn't support metadata directly, but according to this article the route to combine a df and json-like metadata dict in a parquet file doesn't seem too hard. A feature to combine dfs with metadata in our portables is anyways something I am sure that we can benefit from (unless reading the parquet back into pandas becomes a lot slower).

Think this can be a major gain in build time, and possibly also memory usage during build for apps using data from UNSMRY.

[asnyv]

Alternative path to quicker and more memory efficient aggregation of metadata from SMSPEC is to solve equinor/ecl/issues/796 and utilize that in fmu-ensemble / ecl2df with an implementation of the aggregation close to how it currently is (so a separate function for metadata as today in fmu-ensemble)

[anders-kiaer]

Today, from webviz side, we probably want to let the .arrow dump solve efficient metadata aggregation "automatically" (simply be being a good format for arbitrary reads). 🚀