Vectorization can give incorrect results if similar costs depend on values not declared as optim/aux vars

## 🐛 Bug

## Steps to Reproduce

Here is a simple repro

```python
import theseus as th
import torch
from theseus.core.vectorizer import Vectorize

class BadCost(th.CostFunction):
def __init__(self, x, flag=False, name=None):
super().__init__(th.ScaleCostWeight(1.0), name=name)
self.x = x
self.flag = flag
self.register_optim_vars(["x"])

def error(self) -> torch.Tensor:
return (
torch.ones_like(self.x.tensor)
if self.flag
else torch.zeros_like(self.x.tensor)
)

def jacobians(self):
raise NotImplementedError

def dim(self) -> int:
return self.x.dof()

def _copy_impl(self, new_name=None):
return BadCost(self.x.copy(), flag=self.flag, name=new_name) # type: ignore

o = th.Objective()
z = th.Vector(1, name="z")
o.add(BadCost(z, flag=True))
o.add(BadCost(z, flag=False))
o.update() # to resolve batch size for vectorization
Vectorize(o)
print(o.error())

```
## Expected behavior

The above prints `tensor([[1., 1.]])` but it should print `tensor([[1., 0.]])`.