[llvm] [X86][LoopVectorize] Enable MaximizeBandwidth by default on X86 (PR #201666)

[Sumukh J Bharadwaj via llvm-commits]

amd-subharad wrote:

@RKSimon the generic split-cost floor is removed — the change is now just the `shouldMaximizeVectorBandwidth()` override plus regenerated tests. No X86 cost-model change is needed: I audited every VPlan recipe's `computeCost` and each already prices splitting through its TTI hook (`LT.first`/parts), with scalarized recipes multiplying by VF; casts/compares are intentionally sub-linear via the tuned tables, so a floor would over-cost them. Could you take another look and clear the changes-requested when you get a chance?

https://github.com/llvm/llvm-project/pull/201666