Implement gelu (and other elementwise fusions) recomputation during backward

@riccardofelluga is already working on this.

Since Aug 21, 2024.

## 🚀 Feature

### Motivation

```py
import torch
import thunder

def f(x):
x = x @ x
x = torch.nn.functional.gelu(x, approximate="none")
x = x @ x
return x

x = torch.randn(1000, 1000, device="cuda", requires_grad=True)
jf = thunder.jit(f)
y = jf(x)

print([t.shape for t in y.grad_fn.saved_tensors])
# This will print:
# [torch.Size([1000, 1000]), torch.Size([1000, 1000]), torch.Size([1000, 1000])]

print(thunder.core.vjp_utils.get_saved_for_backward_tensors(thunder.last_traces(jf)[-1]))
# This will print:
# (,
# ,
# )
```
In the above snippet `t5` is the output of the gelu function and the request is to implement a pass that forces recomputation of the gelu function in the backward pass instead of saving this intermediate tensor.

Ongoing PR: https://github.com/Lightning-AI/lightning-thunder/pull/1003.

Implementing gelu recomputation would resolve the OOM error seen in https://github.com/Lightning-AI/lightning-thunder/issues/246.