[Feature] Native SIMD Support & MLIR Vector Dialect Integration

We need to implement native SIMD support in the Vx compiler to accelerate AI workloads like the LLaMA-2 inference benchmark.

Proposed Changes:

1. AST & Parser:

   • Introduce Type::Simd(ElementType, usize) into the ast::Type enumeration.
   • Implement syntax for vector types: <N x T> (e.g., <4 x f32>).

2. MLIR Codegen:

   • Lower Type::Simd directly to MLIR's VectorType (vector<4xf32>).
   • Implement lowering for arithmetic binary operations (Add, Sub, Mul, Div) into vector.addf, vector.mulf, etc.
   • Support vector loads and stores via pointer dereferencing.

3. Benchmarks:

   • Modify the transformer() inference loop in llama2_100.vx and llama2.vx to leverage <4 x f32> SIMD operations, replacing purely scalar loops.

This is a critical requirement to remove the $O(N^2)$ scalar math bottleneck in self-attention.

Start by locating the AST/parser and MLIR codegen entry points, then inspect the transformer() inference loops in llama2_100.vx and llama2.vx. The work is done when types, vector arithmetic, loads, and stores lower correctly and the benchmark loops use <4 x f32> SIMD operations.