[llvm] DAG: Consider __sincos_stret when deciding to form fsincos (PR #165169)

[Benjamin Maxwell via llvm-commits]

MacDue wrote:

> Yes, AArch64 should do the same. It happens AArch64 seems to have better test coverage, and I see a regression in the v3 case where the stack slot isn't reused for each element or something

The regression is that rather than expanding early on (and creating three calls to `sincos`), the `v3` case is now marked as legal, but it then gets widened to a `v4`, resulting in another call/stack slot.

https://github.com/llvm/llvm-project/pull/165169