Note: This will currently change how we write rows with adaptive encoding columns. Rows written on older versions of Dolt can still be read, but writing them back will result in a new chunk hash. But since Dolt currently doesn't use adaptive encoding and Doltgres is in beta, that's acceptable.

Claude's Summary

• BuildPermissive now sorts adaptive encoding columns by size savings (largest first) before deciding which to move out-of-band, instead of visiting columns in sequential order
• This keeps smaller values inline when possible, producing shorter tuples when columns have varying sizes
• Added two new test cases verifying the largest-first ordering with 2 and 3 columns of different sizes

Test plan

• Existing TestTupleBuilderAdaptiveEncodings tests pass (including inline one of two columns)
• New test largest value moved out-of-band first: two columns where column 0 is small and column 1 is large — verifies column 1 goes out-of-band while column 0 stays inline
• New test three columns largest first ordering: three columns (medium, small, large) — verifies only the largest goes out-of-band
• Full store/val package tests pass