Fix CUDA MoE router hardcoded to 256 experts

[slackarea]

The CUDA MoE router rejects any model whose routed-expert count isn't 256:
ds4_gpu_router_select_tensor / _batch_tensor do
if (n_expert != 256u || n_expert_used != 6u || fabsf(scale-1.5f)>1e-6f) return 0;
and the three router_select kernels hardcode 256 (logits/probs stride, top-k bound) and
the 1.5f routed-weight scale. On a DeepSeek-V4-Pro GGUF (384 experts) prefill fails with
gpu layer N ffn batch encode failed (the router-select call returns 0).

Fix (minimal, zero-regression for Flash):

• Parametrize the serial router_select_kernel with n_expert and scale.
• Relax both host-wrapper guards to accept 256 or 384; replace the 256u*sizeof(float)
  bias/logits/probs checks with n_expert.
• Dispatch: n_expert != 256 -> the parametrized serial kernel; 256 stays on the
  existing fast warp/parallel kernels (unchanged).

Regression (Linux, CUDA, H200, DeepSeek-V4-Flash q2-imatrix; clean main + this patch only):

• ds4_test --logprob-vectors: PASS
• Flash decode: 39.67 t/s (256 fast path untouched; baseline ~39 t/s)

Caveat / follow-up: still assumes n_expert_used == 6 (the for j<6 loops + guard) — fine
for Flash/Pro (top-6), not for other top-k (e.g. GLM-5.2 top-8). The serial path is
correctness-first (1 thread/token); the warp/parallel kernels could be parametrized for
speed (shared sprob[256] and the 32x8 unroll). Happy to extend if useful.