Enable loop sinking of loads

Enable loop sinking of loads when it's beneficial for register pressure
reduction and rework CodeSinking pass

CodeSinking pass rework:
- Make loop sinking multi-pass to ensure all the beneficial instructions
are sinked
- Use Uniform and AA information
- Add options to force loop sinking for perf tuning
- Disable multi-level loop-sinking (across 2 loop bounds)
- Remove flag-register pressure reduction heuristics
- Add naive load scheduling in local sink
- Make heuristic to decide if the instruction is beneficial to sink aware of the data size
- Change the code style to LLVM in the most parts of the changed functions