No need for two, use the already modernized `docx.shared.lazyproperty`.

Use XPath rather than oxml structures.

This property was formerly known as `._grid_col` but that didn't account
for `.grid_before` in the computation.

This method was formerly named `.tc_at_grid_col()`. New implementation
takes `CT_Tr.grid_before` into account.

The original implementation of `_Row.cells` did not take into account
the fact that rows could include unoccupied grid cells at the beginning
and/or end of the row.

This "advanced" feature of tables is sometimes used by the Word table
layout algorithm when the user does not carefully align the right
boundary of cells during resizing, so while quite unusual to be used on
purpose, this arises with some frequency in human-authored documents in
the wild.

The prior implementation of `_Row.cells` assumed that `Table.cells()`
was uniform and the cells for a row could be reliably be computed from
the table column-count and row and column offsets. That assumption
does not always hold and can raise `IndexError` when omitted cells are
present.

This reimplementation remedies that situation. As a side-effect it
should also perform much better when reading large tables.

`lxml` won't install on Apple Silicon after `4.9.2`.

Dropping testing for Python 3.7.

`.rel_ref_count()` as implemented was only applicable to `XmlPart` where
references to a related part could be present in the XML. Longer term it
probably makes sense to override `Part.drop_rel()` in `XmlPart` and not
have a `_rel_ref_count()` method in `part` at all, but this works and is
less potentially disruptive.

Looks like this cure is worse than the disease.

While it may ease installation on Apple Silicon in some instances, it
breaks installation on Python 3.12. Pull this pin and we'll just have to
live with troublesome `lxml` install on certain Mac/version
combinations.