fix(perf): skip solar-open layer eval during decode

[inureyes]

Summary

SolarOpenModel::forward was calling mlxcel_core::eval_all on the hidden state after every transformer layer to keep the MoE compute graph bounded (128 experts × ~50 ops/layer). That guard is needed for multi-token prefill — without it, the graph grows quadratically with seq_len × n_layers — but it's pure overhead on single-token decode, where the final-logits evaluation flushes the graph once anyway. The per-layer flush was costing ~48 GPU synchronizations per generated token.

Cherry-picked from mlxcel-internal commit c5e88612 (the internal docs/model_tests_m5max.md update was intentionally excluded, same as #20 — it references an internal-only baseline doc that was never in the public repo).

What changed

src/models/solar_open.rs — gates the per-layer eval_all on prefill only:

let eval_layer_outputs = should_eval_layer_outputs(input_ids);  // false when seq_len == 1
for (i, layer) in self.layers.iter().enumerate() {
    h = layer.forward(&h, &mut caches[i], mask);
    if eval_layer_outputs {
        let ptrs = [h.as_ref().unwrap() as *const MlxArray];
        unsafe { mlxcel_core::eval_all(&ptrs) };
    }
}

should_eval_layer_outputs is a 3-line helper that checks input_ids.shape.last() != Some(1), plus a unit test that pins the decode-vs-prefill behaviour.

Verification

• make verify-fmt — clean
• make verify-clippy (CI-faithful: --all-targets --features metal,accelerate -- -D warnings) — clean in 22s (warm cache)
• make verify-test skipped (15–30 min release-mode run); the upstream commit is already validated in mlxcel-internal against the M5 Max sweep.