perf: optimize row merging

[igorbernstein2]

This PR rewrites the row merging logic to be more correct and improve performance:

• extract row merging logic into its own class to simplify complexity of ReadRows handling
• Use OrderedDict instead of dict() for {family: { qualifier: [] }} data, this should maintain serverside ordering (family in creation order and qualifier in lexiographical).
• define an explicit state machine with states implemented as methods
• add various optimizations like:
  • slots on hot objects to avoid dict lookups
  • avoiding dict lookups for contiguous family and qualifier keys

Overall this improves performance by 20% and in my opinion is a lot more readable

[snippet-bot]

No region tags are edited in this PR.

This comment is generated by snippet-bot.
If you find problems with this result, please file an issue at:
https://github.com/googleapis/repo-automation-bots/issues.
To update this comment, add snippet-bot:force-run label or use the checkbox below:

• Refresh this comment

[Mariatta]

Can you add the documentation of row_merger to the Sphinx documentation?
First create docs/row-merger.rst (see the example content from row-data.rst)

Then add row-merger to the usage.rst line 18.

We should also update the usages of row_data.Cell / row_data.PartialRowData / row_data.InvalidChunk in the documentation to use the row_merger instead.

[dbolduc]

The row merging logic LGTM. I left some nits that you can fix or not.