perf(cpu): MoE decode (n=1) routed experts run on k single-thread Parallel.For tasks - two pool dispatches per layer instead #981

Found while investigating why CPU decode on MiMo-V2.6-Flash BPW2.0 (72 GiB mmap) is 1.1 tok/s vs llama.cpp 7.5.

Per-op breakdown (DOTLLM_CPU_PROF=1, warm page cache): ~90% of a decode token is the routed-expert FFN. MoeSwiGluMlp.ExecuteRoutedFromAssignments runs the k selected experts as k independent single-threaded tasks (Parallel.For, inner GEMMs without a pool), serial gate -> up -> down per expert, with ArrayPool rentals per task. Only k cores work, the ComputeThreadPool (spinning) is idle, and a .NET thread pool wake-up is paid per layer.

Fix: for seqLen==1 with a pool, flatten (expert x projection x row) and run two ComputeThreadPool dispatches per layer (gate+up, then down) through the same ComputeRows kernels the n=1 GEMM column uses -> bit-identical to the grouped path.

Acceptance: bit-identical unit test vs grouped path (IQ/K-quant/Q8_0), A/B tok/s per MoE model, env switch DOTLLM_CPU_MOE_DECODE_N1=0.