[llvm] [VPlan] Don't apply predication discount to non-originally-predicated blocks (PR #160449)

[Luke Lau via llvm-commits]

lukel97 wrote:

> Do you by any chance have any perf numbers to confirm this doesn't cause any regressions?
> 
> If so, would be great if you could add them to the PR description

I tested this again on llvm-test-suite arm64-apple-darwin -O3, the geomean difference in loops vectorized is -0.0%. Only four tests were affected with 1 less loop vectorized, and there was no observable performance difference:

```
  MultiSourc...Benchmarks/7zip/7zip-benchmark   355.00                         354.00   -0.3%
  MultiSourc...e/Applications/ClamAV/clamscan   115.00                         114.00   -0.9%
  MultiSourc...ch/consumer-jpeg/consumer-jpeg    65.00                          64.00   -1.5%
  MultiSourc.../mediabench/jpeg/jpeg-6a/cjpeg    61.00                          60.00   -1.6%
  SingleSour...chmarks/BenchmarkGame/fannkuch     1.00                           0.00 -100.0%
```

https://github.com/llvm/llvm-project/pull/160449