Add compact in-memory representation of RecordRef

[Jesse-Bakker]

This adds a compact RecordRef representation, significantly reducing the
per-field overhead for records. There are some more improvements that can be
made here, namely not storing the schema per-record and removing the extra pointer
indirection for the field data.

Overhead

version	Overhead / record	Overhead / field
main	1 fat pointer (16 bytes)	8-24 bytes
compact-record	1 fat pointer + 1 thin pointer (24 bytes)	1 byte

Memory measurements

5M records, 4 integers, 1 float (near-ideal case)

version	RSS(MB)	improvement
main	1390	0%
compact-record	1066	23.3%

5M records, 1 integers, 2 strings

version	RSS(MB)	improvement
main	1508	0%
compact-record	1391	7.7%

[chubei]

Clever use of the fact that Field alignment is less than Field size!

Unfortunately there are some changes we need to revert.

I'm far from proficient in such memory tricks so also left some questions here, mostly to help me understand the implementation.

[chubei]

Wonderful!