Prepare clean Kimi LoRA release tooling

[yyj6666667]

Summary

Prepare a source-built Kimi LoRA release without checkout-only tools or private runtime patches.

• Package the existing LoRA converter and deterministic Kimi/NekoQA preprocessing as CLI tools; retain source-script entry points.
• Document packing-disabled training and full optimizer checkpoint resume.
• Audit all six KT CPU variants separately from CUDA architecture coverage, validate top-level wheel metadata, and use all five release carriers.
• Add hash-checked, dependency-only PEFT/TRL repackaging; runtime files and dependency bounds remain unchanged.
• Preserve wheel-relative dynamic-library paths in the lazy CUDA archive; require matching loader support from SGLang main.
• Round Kimi RAWINT4 scratch allocations to 64 bytes, preventing misaligned access with odd expert counts.

No optimizer or training math is changed; the native change is confined to Kimi scratch-buffer alignment. LF packaging changes live in a separate user-approved fork, not this PR.

Validation

• Release tooling: 68 CPU tests passed.
• Data preparation and converter: 19 CPU tests passed, including fixed public NekoQA data. Five model-directory integration tests were skipped locally.
• Separate sap4 source probe: 138,850 converted ordinary/expert LoRA tensors exactly match the existing input adapter.
• SGLang CPU probe: real ELF loading reproduces the flattened-layout failure and passes with wheel-relative paths; legacy payload behavior remains covered.
• Source-built KT candidate wheel: 24 numerical tests passed with no skips across AMX / AVX512-BF16 and TP1 / TP2, covering RAWINT4 and INT8 forward, input/router/LoRA gradients, accumulation, scaling and frozen base weights. The unfixed main wheel fails both odd-expert cases and passes both even-expert controls.
• Independent LF standard-profile clean install and KT-profile install/upgrade: 46 dependency/dispatch tests pass in each. All 9,945 prepared train/eval label sequences match the original Kimi template. Tiny non-KT CPU training/resume controls reproduce adapter and optimizer state exactly in both environments; this is not Kimi GPU acceptance.
• Installed converter: 138,850 ordinary/expert tensors exactly match the preserved older adapter; inputs are unchanged. This is not a newly trained adapter or generation acceptance.
• Full CUDA source build passes ABI, dynamic-linkage and SASS audits, retaining SM80/86/89/90/120 support and the separate SM90 object. Five-carrier assembly is complete; each wheel is below 104 MB.
• Two fresh environments install the assembled serving / serving+SFT candidates through normal pip dependency resolution. pip check, native imports and comparison against original wheel bytes pass; adding LF/PEFT/TRL leaves all five core packages unchanged. Exact SHA-locked dependency closures also pass pip dry-run.
• Final assembled combined environment repeats all 46 LF and 24 native numerical tests with zero skips, plus exact tiny non-KT optimizer resume. The actual installed SGLang tokenizer factory preserves native Kimi encoding and accepts/rejects JSON-schema inputs correctly on CPU.
• These are PR-source engineering candidates, not official-main release artifacts. New Kimi GPU training → optimizer resume → SGLang generation, quality and cross-model GPU regressions remain pending. Occupied GPUs are not interrupted.

Draft until these gates pass. This PR does not publish packages. SGLang's Kimi tokenizer compatibility fix is tracked in kvcache-ai/sglang#96.