feat(low_latency): support top-16 routing and hidden-dim 3584 (Kimi-K3)

[samsja]

Problem

The low_latency internode dispatch/combine kernels cap top-k at 11 and only template-specialize on a fixed set of a2a hidden dims. Kimi-K3 (896 experts, top-16, routed-expert hidden 3584) doesn't fit:

• top-16 trips EP_HOST_ASSERT(num_topk <= kNumMaxTopK) (dispatch) and the kNumMaxTopk equivalent (combine).
• hidden 3584 falls through SWITCH_HIDDEN's default → "Unsupported hidden". (This is the a2a hidden dim, not the model hidden_size.)

Change (+4/−2)

• kNumMaxTopK / kNumMaxTopk 11 → 16 (dispatch + combine).
• Add case 3584 to SWITCH_HIDDEN — 3584 satisfies the LL alignment invariants (%128, %256 = 32·kNumElemsPerRead, %512 for the combine kNumSendUnrolls=2 path).

Validation

mini-K3 (2 nodes, EP16, --deepep-mode auto, cuda-graph ON): decode cuda-graph capture passes SWITCH_HIDDEN and serves; greedy A/B vs the naive-a2a baseline agrees token-for-token into the low_latency decode region.

Opened against this fork per the sglang team's request (sglang vendors sgl-project/DeepEP:sgl-deepep); an equivalent PR is open against deepseek-ai/DeepEP (deepseek-ai#731).