Add more aggressive late rescheduling

Add more aggressive late rescheduling phase in CodeLoopSinking pass.
This is a phase that happens after the code loop sinking when it's not
enough, unlike existing "early" rescheduling that aims to avoid sinking when
possible.
Currently this more aggressive sinking putting the instructions in
between DPASes in a long DPAS sequence, but tries to split the sequence
only in points when the subsequent DPASes don't share any common
parameters.

This commit also adds an option to disable max sinking heuristic in the
presence of 2d block reads.

Author: admitric