[SLP] Retune look-ahead scores for constants
Rescale LookAheadHeuristics scores and add ScoreSameConstants /
ScoreConstantScaleFactor so identical constants no longer score the
same as same-opcode instruction matches, fixing suboptimal operand
reordering around repeated constants (shift amounts, splat values).
Also retry a SplitVectorize alt-shuffle build for copyable-element
bundles when scheduling fails, instead of always falling back to a
gather, to fix regressions.
Reviewers: hiraditya, bababuck, RKSimon
Pull Request: https://github.com/llvm/llvm-project/pull/207548
