Branch prediction

[georgwiese]

Adds to PGO that for each PC, we collect statistics of the next PC, and persist it to a next_pcs.json file.

The checked-in file has been generated as:
cargo run --bin powdr_openvm -r prove guest-keccak --input 100 --autoprecompiles 1 --apc-candidates-dir keccak100 --mock

For example, consider this memcpy block. Its last instruction is 2105080. In next_pcs.json, you'll find this entry:

  "2105080": {
    "2105084": 300,   // Exit the loop
    "2105036": 2500   // Jump to the beginning of the current basic block
  },

For example, we could generate a precompile that only works when the block is run twice, which would be more efficient than running the original precompile twice (because it only has to accesses the 8 registers once, among other things).

In this (very artificial) run, there are only 7 PCs where the next PC is not PC is not 100% determined, despite there being 63 basic blocks that are executed at least once.