Fix MOI nonlinear derivative sparsity

[harshangrjn]

Based on discussions with @odow , this PR implements the following:

• Fixes inflated Jacobian and Hessian sparsity caused by repeated variables receiving different data slots during MOI translation.
• Reuses variable slots within nonlinear expressions while preserving SIMD batching.
• Adds bounded grouping for nonlinear objectives to avoid large, slow-compiling kernels.
• For PGLib case with 30k nodes, equality Jacobian entries drop from 1,565,483 to 857,623 and Hessian entries from 3,116,094 to 1,558,802, matching the JuMP model’s derivative structure.
• Adds correctness, sparsity, batching, parameter, and Float32 regression tests.

[odow]

I've done pretty much this in #290. We can chat tomorrow.

[odow]

We should discuss this

[odow]

Nice nice. More tests are good.

I'm also planning on enabling the MOI tests. But that was going to wait until we agreed on the changes to the wrapper.