Expand post-lowering optimization pipeline (inline, CSE, LICM, DCE)

[tetsuo-cpp]

Summary

Grow buildWarpForthPipeline beyond a single canonicalizer pass so user words and lowered stack/memory IR get the GPU-relevant opts they need.

Motivation

Current pipeline (lib/Conversion/Passes.cpp):

1. convert-forth-to-memref
2. convert-forth-to-gpu
3. canonicalizer only
4. NVVM attach → GPU→NVVM → math→LLVM → NVVM→LLVM → binary

User-defined words lower to func.func + func.call. On GPU, small helper words should almost always be inlined. Without CSE/LICM/DCE after stack materialization, redundant address math and dead stack traffic survive into PTX.

Proposed passes (order sketch)

After Forth→GPU (and ideally after stack flattening if/when that lands):

1. Inliner — always-inline (or aggressively inline) private word functions into the kernel
2. CSE
3. Canonicalizer (already present; may run more than once)
4. LICM / loop-invariant code motion where profitable
5. DCE / SCCP as appropriate
6. Then existing NVVM/LLVM lowering

Make the set overridable via pass-manager CLI options for experiments (warpforth-opt already registers MLIR passes).

Acceptance criteria

• Small user words are inlined into main in default pipeline IR dumps
• Documented pass list in README or architecture notes
• LIT tests showing inlining and at least one CSE/LICM win on a reduced example
• No compile-time regression on existing check-warpforth suite
• Pipeline remains debuggable: ability to stop before/after opt stages (see also dump stages on warpforthc)

Related

• Stack flattening (promoted SSA makes these opts much more effective)
• Loop unrolling hints (Loop unrolling hints #12)
• Configurable GPU arch (separate issue)