Consider update structure overhead in split length calculation

For Linkbench with maxid=200M the database size in GB after 1 hour of transactions is:

• 451, 382, and 367 for rc13, rc9 and 4.2.8 without Snappy compression
• 233, 215 and 204 for rc13, rc9 and 4.2.8 with Snappy compression

The regression started with rc12 although I didn't list the results for that above. Much more performance data is here. I don't see a significant regression in size for the insert benchmark. It is ~10% there.

The biggest growth is from _id which is ~2X larger now in rc12 & rc13 compared to 4.2.8. That looks like a return of WT-6251