[llvm] [SLP] Emit reduction instead of 2 extracts + scalar op, when vectorizing operands (PR #147583)

[Alex Bradbury via llvm-commits]

asb wrote:

I've pushed a revert for the time being because this break the RVV builders too (MicroBenchmarks/ImageProcessing/Blur/blur.test and MultiSource/Benchmarks/tramp3d-v4/tramp3d-v4.test from llvm-test-suite)


https://github.com/llvm/llvm-project/pull/147583