Separate P/D logic from the prefix scorer

[kfirtoledo]

Currently, part of the P/D logic is implemented in the prefix scorer. This should be moved to the pd package and use the prefix indexer for hit map calculation.

[kfirtoledo]

Waiting for PR: added cycle state to pick & process results in profile handler to be merged. (cc @nirrozenbaum)
With this PR, we can read in the pick function the cycleState, which contains the prefix score for the selected decode pod, and do the calculation for hitPercentage (hitPercentage := Score*prefixStore.cacheBlockSize/len(prompt)). The hitPercentage is used to decide whether prefill is needed.

[nirrozenbaum]

in the process results (not in pick) 🙂

[nirrozenbaum]

GIE PR was merged. please see #197 that was submitted to sync llm-d-scheduler with latest GIE (the one that includes the required change with CycleState).