Add Kimi-K3 text prefill exporter

[micwill755]

Summary

• add a Kimi-K3 text prefill EdgeSpec for the 0.40B development checkpoint
• export the hybrid KDA/MLA stack with attention-residual and dense MoE patches
• register and lower the kimi_kda custom op with explicit convolution and recurrent state inputs
• add a standalone Kimi exporter smoke script

Validation

• focused pre-commit hooks pass
• Kimi-K3-0.40B eager/TRT parity: 100% close, relative L2 0.0005
• measured 14.110 ms eager vs 1.106 ms TensorRT on the local smoke input

Scope

This PR targets text-only static prefill and is stacked on the Nanbeige exporter PR. The experimental Edge-LLM KDA plugin implementation is being developed separately.