Thunder-generated backward for sigmoid may compute `nan` for large (in absolute values) negative inputs

@kiya00 is already working on this.

Since Oct 9, 2025.

## 🐛 Bug

The derivative of `sigmoid(x)` is `sigmoid(x) * (1 - sigmoid(x))` which is obtained after a couple of algebraic simplifications and substitutions. There's no manual implementation of this derivative in Thunder and therefore it's generated from its decomposition and chain rule that together with fusion rematerialization results in:
```py
def backward(output_grad, x):
return -(-output_grad * sigmoid(x) * sigmoid(x) * exp(-x))
```

Reproducer:
```py
In [1]: import thunder
/home/iyashchuk/dev/lightning-thunder/thunder/executors/transformer_engineex.py:56: UserWarning: transformer_engine failed to import with exception cannot import name 'MXFP8BlockScaling' from 'transformer_engine.common.recipe' (/home/iyashchuk/miniforge3/envs/pytorch-cuda-dev/lib/python3.10/site-packages/transformer_engine-1.11.0+c27ee60-py3.10-linux-x86_64.egg/transformer_engine/common/recipe/__init__.py)
warnings.warn(f"transformer_engine failed to import with exception {ex}")
/home/iyashchuk/dev/lightning-thunder/thunder/executors/transformer_engineex.py:62: UserWarning: Installed version of transformer_engine 1.11.0+c27ee60 is not supported, please upgrade to version 2.0 from https://github.com/NVIDIA/TransformerEngine/tree/release_v2.0. `transformer_engine_ex` will not be used.
warnings.warn(msg)

In [2]: import torch

In [3]: @thunder.jit
...: def sigmoid(x): return torch.sigmoid(x)

In [4]: a = torch.tensor(-1000., device="cuda", requires_grad=True)

In [5]: y = sigmoid(a)

In [6]: y.backward(y)

In [7]: a.grad
Out[7]: tensor(nan, device='cuda:0') # INCORRECT!

In [8]: a.grad = None

In [9]: y = torch.sigmoid(a)

In [10]: y.backward(y)

In [11]: a.grad
Out[11]: tensor(0., device='cuda:0')
```
Here's the backward execution trace:
```py
In [13]: sigmoid._lc_cs.last_backward_traces[-1]
Out[13]:
def backward_fn(saved_for_backward, cotangents):
# saved_for_backward: "Collection"
# cotangents: "Collection"
C0, _, = saved_for_backward
# C0: "Collection"
# None
clear_mutable_collection(saved_for_backward)
del saved_for_backward
t8, = cotangents
# t8: "cuda:0 f32[]"
clear_mutable_collection(cotangents)
del cotangents
t3, x, = C0
# t3: "cuda:0 f32[]"
# x: "cuda:0 f32[]"
clear_mutable_collection(C0)
del C0
[bw_t13] = nvFusion0(t8, x, t3)
# bw_t9 = prims.neg(t8) # bw_t9: "cuda:0 f32[]"
# t4 = prims.neg(x) # t4: "cuda:0 f32[]"
# bw_t10 = prims.mul(bw_t9, t3) # bw_t10: "cuda:0 f32[]"
# t5 = prims.exp(t4) # t5: "cuda:0 f32[]"
# bw_t11 = prims.mul(bw_t10, t3) # bw_t11: "cuda:0 f32[]"
# bw_t12 = prims.mul(bw_t11, t5) # bw_t12: "cuda:0 f32[]"
# bw_t13 = prims.neg(bw_t12) # bw_t13: "cuda:0 f32[]"
del t8, x, t3
return (bw_t13,)
```