[llvm] [SLP]Fix fmul/fadd fusion costs and retry FMA seeds after all blocks (PR #226117)

[via llvm-commits]

bipmis wrote:

Another one observed in a Fortran benchmark. Find the reproducer attached[bwaves-simple.tar.gz](https://github.com/user-attachments/files/32916311/bwaves-simple.tar.gz)

https://github.com/llvm/llvm-project/pull/226117