[llvm] [SLP] Partially revert #224931 for alternate nodes (PR #228659)

[Dmitry Sidorov via llvm-commits]

https://github.com/MrSidims commented:

The scalar context carries more than its user., for example AMDGPU reads the afn flag of an fdiv and X86 reads the strictfp attribute through it and only the fmul price depends on the user.

I would drop the context for the fmul alone and keep it for the other opcodes, the way #228902 does for a vector fmul whose lanes get extracted. That leaves your test unchanged and keeps the fast-math flags and function attributes reaching the target in the alternate node.

https://github.com/llvm/llvm-project/pull/228659