[gvar] we could allocate less when building tuple variations

[cmyr]

Choosing the correct packed points format for a given tuple requires us to fully build that tuple twice, after which we throw away these results; we then build the tuple a third time, at the end. In theory we could have some shared set of buffers that we reuse for all of this work, which would be slightly more complicated.

In practice this is very low priority, but something to look at as directed by profiling.