Add __nv_sincosf and __nv_sincos

[cncastillo]

I added NVIDIA’s sincos symbols to the existing sincos CallPattern:

def : CallPattern<(Op $x),
  ["__fd_sincos_1", "__fd_sincos_1f", "__fd_sincos_1l",
   "__nv_sincos", "__nv_sincosf"],

Fixes #2552.

However, I am not sure this is actually correct, ChatGPT says the ABIs differ (below):

---

Current Enzyme test uses a struct‑return function, e.g.:

declare [2 x double] @__fd_sincos_1(double)

whereas the NVIDIA libdevice intrinsics are pointer‑out, void return:

__device__ void __nv_sincos (double x, double *s, double *c);
__device__ void __nv_sincosf(float  x, float  *s, float  *c);

So with this change I am effectively binding symbols with a [2 x T] (T) ABI and symbols with a void (T, T*, T*) ABI to the same CallPattern. I am worried this might be wrong or at least fragile, since the existing derivative logic for sincos appears to assume a non‑void aggregate return (as in ReverseMode/sincos.ll).

---

I am not really sure how to test this, any pointers would be appreciated.

[wsmoses]

yeah so I think we need to perform a transformation to change it to be struct-return in order to use the tablegen utilities here, we need to define the rule manually in enzyme, or we need to upgrade non-blas tablegen to handle mutation