[feat](kt-kernel): add native MXFP4 routed-expert LoRA SFT

[yyj6666667]

What does this PR do?

This PR adds standalone CPU LoRA forward/backward support for DeepSeek V4 native MXFP4 routed experts.

Main changes:

• Consume the checkpoint-native E2M1 weights and group-32 UE8M0 scales without creating a persistent dense BF16/FP32 base copy.
• Compute base dX directly from packed row-major weights, without an additional transposed base-weight copy.
• Implement the DeepSeek V4 asymmetric SwiGLU clamp and its backward derivative, with LoRA applied before the clamp.
• Reuse KT authoritative LoRA gradients: overwrite on the first microbatch, accumulate afterward, and lazy-clear between optimizer windows.
• Add source-level, compiled numerical, lifecycle, and real-checkpoint layer validation.

The scope of this PR is one CPU routed-expert MoE layer. Whole-model DeepSeek V4 attention backward and training-framework integration are left for a follow-up.

Related to #1921.

Tests

• Python reference/API/gradient lifecycle: 49 passed.
• Compiled TP1/TP2 numerical tests: GAS 1/2/8, different active-expert sets, lazy clear, checkpoint-cache replacement, and legacy MXFP4 inference regression passed.
• Synthetic comparison against an FP32 reference: worst relative L2 0.008501, minimum cosine 0.999965.
• Real DeepSeek-V4-Flash-0731 layer 3 smoke: 256 experts, H=4096, I=2048, top-k=6, rank=8; TP1 and TP2 passed.
• Real-layer comparison: worst relative L2 0.007374, minimum cosine 0.999973; inactive expert gradients remained zero and the packed base checksum was unchanged after backward and a LoRA update.
• CPU-only AVX512-BF16 extension rebuilt successfully on dual AMD EPYC 9355.

Before submitting

• Did you read the contributor guideline?
• Did you write any new necessary tests?