Speed up performance for DataRows

[mpolson64]

Summary:
Misc improvements and tricks to make DataRows more performant. We're within spitting distance of the original implementation with dataframes, close enough that Im willing to consider the difference is likely due to scheduler noise; IMO good enough to land.

1. Removed isinstance check from Data init -- this was helpful when refactoring since some calls to Data(df) didnt use kwargs and caused errors, but added unnecessary overhead
2. [BIG IMPROVEMENT] Used df.itertuples instead of df.iterrows in Data init when initializing from a dataframe. This alone took us from 1h 44m to ~40m
3. New empty, metric_names, and trial_indices properties which dont require constructing full_df
4. Changes to Experiment.attach_data which operate directly on list[DataRows] instead of on DataFrames (ie migrating from combine_df_favoring_recent helper fn to new combine_data_rows_favoring_recent fn)
5. Changed [*foo] to list(foo) in a couple places. Metamate tells me this is faster in extremely high data regimes -- not sure I notice a difference or trust it necessarily.

Remaining TODOs:
Id be interested in removing property from the methods which are not O(1); theres a lot of fairly expensive things we do in Data, or at least things which require a full scan, which look like they should be fast because they have the same syntax as an attribute lookup. If nobody has any objections to this Ill ask Metamate to do this for us

Differential Revision: D90713603

[meta-codesync]

@mpolson64 has exported this pull request. If you are a Meta employee, you can view the originating Diff in D90713603.