Split `std::batching` out of `std::autodiff` to enable AoS and SoA vectorization.

The std::autodiff feature allows batching of calls. So instead of calling the function f(x: f32) -> f32 four times with four different arguments, one can call batched_f(x1: f32, x2: f32, x3: f32, x4: f32) -> [f32;4].

In the Python world, this feature is known from JAX as .vmap().
std::batching should allow Array-of-struct and Struct-of-array vectorization. It is a common autodiff feature since e.g. ML frameworks allow batching to improve training performance, but we have enough interested users who are not interested in autodiff/derivatives/gradients, so it would be good to also offer it as standalone feature.

The first step is to add a std::batching intrinsic, and lower it to an __enzyme_batch call on llvm-ir level.
Luckily, this has been done for autodiff as part of the latest GSOC project, so you can copy most of it: https://github.com/rust-lang/rust/pull/142640. In this first MVP part, you just need to focus about getting the infrastructure up, even handling a single testcase is enough. The parsing logic will be a strict subset of the parsing logic for std::autodiff, so it should be straightforward.

In a second PR we should then talk with @ Jubilee and some other people from the std::simd project to improve the actual types which we generate. For example, it is likely that we'd want to generate Rust's SIMD types in most cases instead of just generating arrays. Enzyme has a lot of configurations, and we should talk about how to map them nicely to Rust types. When implementing std::autodiff I exposed these in one way, but we likely want to re-evaluate my choices. Any improvements can later be fed back into the std::autodiff macro.

Please ask any questions you might have here and ping @Sa4dUs for questions about introducing an intrinsic, and me about the Enzyme/autodiff/batching logic or lowering to LLVM.

Start by reading the linked Rust pull request and the existing std::autodiff intrinsic, parsing, and LLVM lowering implementation. Trace how the autodiff path reaches LLVM IR, then scope the MVP to a std::batching intrinsic that lowers to __enzyme_batch and supports one testcase. Done means the infrastructure builds and that testcase passes; SIMD type design is explicitly deferred.