Re-apply "[AMDGPU][Scheduler] Scoring system for rematerializations (#175050)" (#177206)

This re-applies commit f21e3593371c049380f056a539a1601a843df558 along
with the compile fix failure introduced in
8ab79377740789f6a34fc6f04ee321a39ab73724 before the initial patch was
reverted. It also fixes for the previously observed assert failure.

We were hitting the assert in the HIP Blender due to a combination of
two issues that could happen when rematerializations are being rolled
back.

1. Small changes in slots indices (while preserving instruction order)
compared to the pre-re-scheduling state means that we have to re-compute
live ranges for all register operands of rolled back rematerializations.
This was not being done before.
2. Re-scheduling can move registers that were rematerialized at
arbitrary positions in their respective regions while their opcode is
set to DBG_VALUE, even before their read operands are defined. This
makes re-scheduling reverts mandatory before rolling back
rematerializations, as otherwise def-use chains may be broken. The
original patch did not guarantee that, but previous refactoring of the
rollback/revert logic for the rematerialization stage now ensures that
reverts always precede rollbacks.