How to do freeze VarMap Vars?

Hello everybody,

Is there away to freeze all Var Tensors in the VarMap like the below snippet ?
means something like implement the `Iterator` trait and detach the contained tensors from the graph and add a Var which can be trained !!!

```
# Freeze all the pre-trained layers
for param in model.parameters():
param.requires_grad = False

```

_Originally posted by @mohamed-180 in https://github.com/huggingface/candle/issues/891#issuecomment-2214407719_

Start by reading the VarMap and Var APIs and checking how Var tensors participate in the computation graph; the issue names no files or tests. Define how freezing all existing variables should coexist with adding a trainable Var, then verify the behavior with a focused test for frozen pretrained layers and the new trainable variable.

Written by the indexing model from the issue text.