[tetsuo-cpp]

Why

warpforthc always compiled a Forth kernel through to PTX. Inspecting an earlier stage meant splitting the compile across warpforth-translate and warpforth-opt. This adds one flag so the same driver can stop after parsing, after lowering, or at PTX.

What changed

• warpforthc --emit=mlir|llvm|ptx writes the parsed Forth module, the lowered module as MLIR, or PTX. The default is ptx.
• buildWarpForthLoweringPipeline holds the stages before PTX emission. The full pipeline calls it, then emits PTX as before.

Scope

• Covers the remaining stage-stop flag from issue 120.
• Leaves out real LLVM IR text and a compilation-target flag on this driver.

Tradeoffs

• --emit=llvm writes the lowered module as MLIR text. A .ll export would add a second translator and a new link dependency for a debug stop.

Blast Radius

• A run without --emit still writes PTX. That output matched the canon binary on the new test kernel. warpforth-opt is unchanged.

Verification

• Ran warpforthc --emit=mlir, --emit=llvm, --emit=ptx, and a default run on test/Pipeline/warpforthc-emit.forth. The default run and --emit=ptx matched each other and the canon binary. --emit=cubin exited 1 with Cannot find option named 'cubin'!.
• The worktree lit suite passed 138 of 138, including the new emit test and the existing options test.

Closes #120