Non-power-of-2 SLP vectorization regressions (and improvements) on RISC-V

LNT picked up a handful of regressions after the non-power-of-two SLP vectorization change #196825 : http://lnt.llvm.org/db_default/v4/nts/22

Just flagging this here as there's likely some cost modelling issues to be worked through, see

- [SingleSource/Benchmarks/BenchmarkGame/n-body](http://lnt.llvm.org/db_default/v4/nts/22/graph?test.1339=3)
- [MultiSource/Benchmarks/McCat/08-main/main](http://lnt.llvm.org/db_default/v4/nts/22/graph?test.1224=3)
- [MultiSource/Benchmarks/TSVC/LinearDependence-flt/LinearDependence-flt](http://lnt.llvm.org/db_default/v4/nts/22/graph?test.1290=3)

There are also some nice improvements too though, [SingleSource/Benchmarks/Shootout-C++/Shootout-C++-matrix](http://lnt.llvm.org/db_default/v4/nts/22/graph?test.1428=3) is 12% faster