Optimization presets: implement -O0, -O1, -O2, -O3 pass pipelines

[yudongusa]

Background

Currently the `PassManager` is a flat, user-assembled sequence. Production compilers expose standard optimization levels (-O0 through -O3) that choose a well-tuned sequence of passes. Users and frontends need this — nobody wants to manually compose 20 passes.

Proposed API

// llvm-transforms/src/pipeline.rs
pub enum OptLevel { O0, O1, O2, O3 }

pub fn build_pipeline(level: OptLevel) -> PassManager {
    let mut pm = PassManager::new();
    match level {
        OptLevel::O0 => { /* no passes */ }
        OptLevel::O1 => {
            pm.add_function_pass(Mem2Reg);
            pm.add_function_pass(DeadCodeElim);
            pm.add_function_pass(ConstProp);
        }
        OptLevel::O2 => {
            // O1 passes +
            pm.add_function_pass(Inliner::default());
            pm.add_function_pass(Gvn);
            pm.add_function_pass(LoopUnroll::default());
        }
        OptLevel::O3 => {
            // O2 passes +
            pm.add_function_pass(LoopVectorize::default());
            pm.add_module_pass(Ipa);
        }
    }
    pm
}

Required new passes for O2/O3

• GVN (Global Value Numbering) — see issue Advanced middle-end passes: GVN, loop unrolling, and function inlining improvements #85
• Loop Unrolling — see issue Advanced middle-end passes: GVN, loop unrolling, and function inlining improvements #85
• SIMD Vectorization — see issue SIMD auto-vectorization: emit SSE4/AVX2 instructions from vector IR types #86
• IPA (Inter-Procedural Analysis) — see issue Inter-procedural analysis (IPA): whole-program optimization across function boundaries #87

Acceptance criteria

• build_pipeline(OptLevel) function in llvm-transforms
• O0 and O1 pipelines working with current passes
• Integration test: same .ll file run through O0 vs O2 produces semantically equivalent but measurably smaller IR
• Benchmark showing wall-clock speedup of O2 vs O0 on the sample.ll fixture
• Exposed in the examples/tikv_jit example via a --opt-level flag

References

• LLVM's default pass pipelines