moe: correct + deterministic KV-cached gpt-oss decode (single card)

Root-caused the decode non-determinism: matmul's m==1 custom-GEMV fast path
reduces over K with a grid-split atomicAdd, whose float accumulation order
is non-deterministic. Negligible for attention's stable pre-transposed
weights, but for gpt-oss's wide expert GEMMs (K=2880, N up to 5760) over
freshly-dequantized MXFP4 weights it produced visibly different results
run-to-run (and a wrong argmax). Added gemm::matmul_dense (plain cublasGemmEx,
no GEMV shortcut) and route the expert GEMMs through it.

Now decode_step (KV cache + GPU sink-attention + MXFP4 experts) is:
- deterministic: 3/3 identical runs
- correct: top-1 token 12650 = " Paris" for "The capital of France is",
  MATCH_TOP1 with the host-attention reference forward
- end-to-end: gptoss-gen generates 32 tokens at ~6.85 tok/s on one 5090.

Removed the temporary A/B debug dumps. gptoss-logits runs both paths and
asserts the top-1 match; gptoss-gen times greedy generation.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

crates/xserv-model/src/bin/gptoss-logits.rs

@@ -35,10 +35,7 @@ fn main() {
     }
 
     // (2) KV-cache GPU decode path (token-by-token prefill) — must match top-1.
-    let mut cache = xserv_model::GpuKVCache::new(
-        model.config.num_layers(), model.config.num_kv_heads(),
-        model.config.head_dim(), xserv_tensor::DType::BF16, Device::Cuda(0),
-    );
+    let mut cache = xserv_model::GpuKVCache::new(&model.config, 512, xserv_tensor::DType::BF16, 0);
     let mut dlog = model.decode_step(tokens[0], &mut cache);
     for &tok in &tokens[1..] {
         dlog = model.decode_step(tok, &mut cache);