LFM2.5-24B-A2B on CUDA: plain serve on top of #1377 #1488

LFM2.5-24B-A2B (lfm2_moe) runs natively on Metal through BlivionIaG's stack (#1377, DSpark in #1390–#1394). I'm porting the plain serve to CUDA, on top of #1377, as discussed with BlivionIaG.

Scope of the first PR: plain serve (no drafter) of LiquidAI/LFM2.5-24B-A2B-MLX-4bit on CUDA. It reuses the shared CUDA pieces wherever they fit: norms, the 4-bit affine GEMV/GEMM tiles, MoE glue and paged attention. The head-64 paged attention segments move into a neutral paged_attn_segs.cuh, with Nemotron's symbols unchanged. What's new is the short-conv layer with its state, its commit, and the router (fp32 weights, expert bias, normalized top-4).

Gates: the kernel contract for the new kernels, SASS identity for everything shared, the float64 reference, windowed vs one-row decode, concurrent vs solo, and a served receipt on GB10.

Later, separately: the DSpark drafter on CUDA, after #1390–#1394 land.