[MODEL] support Qwen3.8 Flash Next quantization

[ZX-ModelCloud]

Summary

• add Qwen3.8 Flash Next / qwen4_exp quantization support for sparse attention, linear attention, shared experts, and routed experts
• reuse the qwen4_exp_text checkpoint conversion mapping and materialize sharded PLE weights on CPU
• preserve CPU-resident PLE during calibration replay and quantized checkpoint loading to avoid GPU OOM
• add model registration, eval score coverage, focused unit tests, and README documentation

Tests

• focused Qwen4-Exp, PLE lazy-load, and forward-placement tests: 10 passed
• py_compile on all changed Python files
• ruff check on the changed files, excluding pre-existing repository-wide E402, E701, and F541 findings

Depends on ModelCloud/Defuser#64.