feat: support GLM-5.2

[CUHKSZzxy]

Motivation

Add GLM-5.2 PyTorch serving support and improve FP8 decode efficiency and stability. The DSA optimizations are shared with DeepSeek-V3.2 where applicable.

Changes

• Add GLM-5.2 model/config integration, DSA indexing and cache handling, tool-call parsing, expert/indexer replay, and MoE-only FP8 support.
• Replace the memory-heavy bitonic sparse top-k path with radix top-k, avoiding a score-sized temporary index buffer.
• Fuse DSA indexer preparation: remove redundant Hadamard transforms; fuse Q RoPE, FP8 quantization, and gate scaling; fuse K LayerNorm, RoPE, FP8 quantization, and cache writes; and merge K/head-gate projections.
• Support both GLM interleaved RoPE and DeepSeek-V3.2 NeoX RoPE. LMDEPLOY_DISABLE_DSA_INDEXER_FUSION=1 preserves the original path.
• Use native Transformers configs where available, with local fallback registration for DeepSeek-V3.2.

Validation

• Pre-commit passed; 25 config/model tests and 8 CUDA indexer tests passed.
• Real 5-layer FP8 smoke tests passed for GLM-5.2 and DeepSeek-V3.2.
• Short batch-8 decode timelines reduced median cycle time by 3.41% for GLM-5.2 and 4.66% for DeepSeek-V3.2, removing 13 CUDA graph child kernels per DSA layer. These are profiling results, not full-model throughput claims.

Assistance

Assisted with Codex + GPT-5.6-Sol xHigh, reviewed manually