Add Kimi K2 target integration path

[manishklach]

What changed

• introduces a strict, fail-closed Kimi K2 Thinking checkpoint configuration and Safetensors inventory adapter
• adds immutable manifests and deterministic stationary-owner plans before any GPU allocation
• adds a paged-KV decode semantic oracle, bounded continuous batching, telemetry, and resilient transport behavior
• adds a fused packed-INT4 SwiGLU HIP correctness ABI and documents its explicit performance boundary

Why

Kimi K2 Thinking is an official open-weight trillion-parameter sparse MoE target with documented MLA/SwiGLU semantics. The project can now admit only checkpoint artifacts that match the target contract while keeping hardware-dependent compatibility certification gated on actual ROCm and real-checkpoint fixtures.

Validation

• python -m pytest -q — 18 passed
• python -m ruff check . — passed
• python -m build — source distribution and wheel built

Non-claims

This does not claim Kimi K2 compatibility, MFMA performance, or GPUDirect RoCE. The release gates require target hardware, official checkpoint fixtures, and parity/benchmark measurements before those labels are allowed.