EP: support Kimi-K3 (top-16 routing and hidden 3584) (#2080)

Kimi-K3 routes top-16 experts and dispatches a 3584 latent-MoE hidden
size; both are rejected by the current low-latency limits.

- nixl_ep_ll.cu: raise the dispatch and combine top-k caps from 11 to 16
- launch.cuh: add hidden 3584 to SWITCH_HIDDEN

<!-- This is an auto-generated comment: release notes by coderabbit.ai
-->

## Summary by CodeRabbit

* **New Features**
  * Added support for hidden size `3584` configurations.
* Increased the maximum supported top-k value from 11 to 16 for dispatch
and combine operations.

<!-- end of auto-generated comment: release notes by coderabbit.ai -->

Signed-off-by: Lior Shapira <lishapira@nvidia.com>