feat(smollm2): compiled-export module — redecode graph with argMax tail (#305) - #308

Summary

New :llm-inference:smollm2 module — the whisper/functiongemma "one export
surface" shape. SmolLm2ExportHarness.export() traces llamaNetwork() from
the real SmolLM2 GGUF, strips per-layer KV caches, and appends the DSL argMax
tail so the emitted func @smollm2 returns tensor<seqxi32> directly
(small per-step output — no host-side argmax over [seq, vocab] logits).
This is the redecode-graph counterpart to FunctionGemmaExportHarness's
exportRedecode and matches GemmaDecoder's consumption pattern: one
fixed-seq vmfb invoked in a loop over a growing, causally-masked-safe padded
buffer (the two-graph KV-cache decode is deliberately out of scope — see #305).

Depends on :llm-inference:llama for the architecture; owns the export
product, mirroring functiongemma's relationship to gemma.

Verification — the full pipeline, numerically, not just compiled

1. SmolLm2ExportHarness.export() against the real
   SmolLM2-135M-Instruct-Q8_0.gguf: 393 params, 310 MiB bf16 archive,
   func.func @smollm2(%arg0: tensor<1x24xi32>) -> (tensor<24xi32>).
2. iree-convert-parameters → smollm2.irpa, iree-compile --iree-hal-target-backends=llvm-cpu → a host vmfb (via
   SKaiNET-iree-toolchain's skainet/iree-compiler image).
3. Drove the GemmaDecoder-style redecode loop by hand via
   iree-run-module for the prompt "The capital of France is" (tokens
   1,504,3575,282,4649,314): 8 greedy steps decode to "the city of
   Paris, a city of" — correct and coherent. This confirms the DSL trace,
   argMax tail, bf16 externalization, and IREE execution are numerically
   sound end to end, not just that the graph compiles.

What's not in this PR

Per #305's split: compiling this graph for Android arm64 and the JNI decode
module that drives it on-device. I have a working (compiled, symbol-verified)
spike of that JNI shim over the IREE C API — cross-built via
SKaiNET-iree-toolchain's skainet/iree-android image — but it isn't
landed anywhere yet pending a decision on where it should live (new repo
vs. existing).

Test plan

• ./gradlew :llm-inference:smollm2:compileTestKotlinJvm — clean, no warnings
• SMOLLM2_GGUF=... ./gradlew :llm-inference:smollm2:jvmTest --tests '*SmolLm2ExportHarnessTest*' -PincludeIntegration — passes (393 params, 310 MiB, correct function signature)
• Manual end-to-end numeric verification via iree-compile + iree-run-module (see above)
• CI (no model file present, so the integration test self-skips — expected)

Advances #305.