Some potential optimization points

[forsaken628]

Record some fragmented thoughts

[forsaken628]

When sorting on a single nullable simple type, the sort row is also encoded as a binary row. And the index of a binary row requires two buffer accesses, which is the same as a nullable column, so this encoding is redundant and can be removed.

[forsaken628]

sort spilling does not always require serialization and deserialization of the extra generated sort row, which helps reduce io.

[forsaken628]

Current sort rows mainly use binary encoding, which is a variable-length encoding. Fixed-length encoding needs to be supported.