Fix problems in TimeSeries primary key handling - #20361

Always give a TimeSeries a primary index on time, whatever indices its input columns carry

1. Why the primary key ends up None

Table.primary_key is a plain attribute that only Table.add_index() sets, and only when the table has no indices yet. Column copies made by col_copy(..., copy_indices=True) deep-copy the column's SlicedIndex objects, so a table built from such columns has indices but no primary key. Three paths in astropy.timeseries hit this:

• ts["time", "a"] goes through Table.__getitem__, which builds a new TimeSeries from the copied columns. In TimeSeries.__init__ the time column is then removed and re-added, but it is the same column object, still carrying its copied index. add_column sees len(self.indices) == 1, skips add_index("time"), and primary_key is never set.
• TimeSeries(time=ts.time) copies the indexed Time column in add_column, with the same outcome. This is TimeSeries.iloc failed when time comes from another TimeSeries #11704.
• BinnedTimeSeries.__init__ has the same remove/re-add pattern for time_bin_start and then calls add_index("time_bin_start") unconditionally, so the sliced series has two identical indices, neither of them primary.

TableLoc._get_index_id_and_item then returns self.table.primary_key, i.e. None, and TableIndices.__getitem__(None) falls through to list.__getitem__, giving the TypeError in the report.

A related symptom on main: ts.fold() on a series with a second index (say ts.add_index("a")) removes and re-adds time, and the len(self.indices) == 0 guard then leaves the folded series with a primary key of ("time",) but no index on time.

2. Implementation

astropy/timeseries/core.py, new BaseTimeSeries._add_primary_index(colname):

if (colname,) not in [index.id for index in self.indices]:
    self.add_index(colname)
self.primary_key = (colname,)

It adds an index on the column only if there is none, and sets the primary key explicitly because add_index does that only for the first index of a table.

astropy/timeseries/sampled.py, add_column / add_columns: the len(self.indices) == 0 and "time" in self.colnames guard becomes if "time" in self.colnames: self._add_primary_index("time"). The constructor's remove/re-add of time goes through add_column, so this single change covers the column-slice case, #11704, and fold().

astropy/timeseries/binned.py, BinnedTimeSeries.__init__: add_index("time_bin_start") becomes _add_primary_index("time_bin_start"). That removes the duplicate index for both the column-slice case and BinnedTimeSeries(time_bin_start=other.time_bin_start, ...).

Nothing was removed from the public API. The approach in #20354 (an elif self.primary_key is None: self.primary_key = ("time",) in add_column) is not used because it assigns a key without checking that an index on time exists; with qt.add_index("a"); TimeSeries([qt["time"], qt["a"]]) it produced IndexError: No index found for ['time'] instead.

3. Backwards compatibility

No public API changes. A TimeSeries whose time column has no carried-in index behaves exactly as before: add_index("time") is called once and becomes the primary key. What changes is only the previously broken states: column-sliced or column-constructed series now have a usable primary time index, BinnedTimeSeries no longer accumulates duplicate time_bin_start indices, and fold() on a series with a second index keeps its time index. A series that carries a second index (e.g. via ts.add_index("a")) always has time as the primary key after any add_column; before, the primary key was whichever index was added first, which for a TimeSeries was already time.

Changelog fragment: docs/changes/timeseries/20361.bugfix.rst.