cache stems in case multiple values are used

[gballet]

In order to reduce the amount of double computations, one should cache stems when they are computed. This way, they can be reused:

• when writing two slots in the same group when updating the tree
• for any write/read that has to happen before the final tree update at the end (FILL_COST?)
• for the prefetcher later on

This might not be as interesting as computing stems in one swoop but it's worth checking if it positively impacts performance.