Differentiation of conic programs with nonlinear parametric expressions

DiffOpt does not support differentiating a conic program if the problem's data depends nonlinearly from the parameters (e.g. if `A(theta) x <= b(theta)` with `A(.)` and `b(.)` given by nonlinear expressions).

Here is a simple MWE:
```julia
using JuMP
using Clarabel
using DiffOpt

model = DiffOpt.diff_model(Clarabel.Optimizer)
@variable(model, x >= 0)
@variable(model, θ ∈ MOI.Parameter(1.0))
@objective(model, Min, x)
@constraint(model, sin(θ) * x >= 1.0)
JuMP.optimize!(model)

```

If we replace Clarabel by Ipopt, the code works like a charm. But Clarabel works much better than Ipopt if the problem is conic.

We can use the chain-rule to differentiate the problem w.r.t. `A` instead of `theta`. However, I think it would be more intuitive to support nonlinear expressions depending on the parameters in DiffOpt. What do you think?

Start by running the Julia MWE with Clarabel and compare it with the Ipopt version, then read the DiffOpt parameter-differentiation path reached through diff_model and optimize!. Done means nonlinear parameter expressions such as sin(θ) * x are supported for conic programs and the supplied example differentiates successfully.

由索引模型根据 Issue 内容生成。