[llvm] [SLP]Flatten fsub/fneg chains into fadd reductions with per-operand signs (PR #215840)

[Tomas Matheson via llvm-commits]

tommat01 wrote:

As mentioned in #221717 this seems to cause a regression in graviton4 spec/lbm. 
[lbm-lto-signed-reduction.ll.txt](https://github.com/user-attachments/files/32565341/lbm-lto-signed-reduction.ll.txt)

The code is partially vectorised because the issue originally surfaced in LTO, after pre-LTO vectorisation had run.

https://github.com/llvm/llvm-project/pull/215840