feat(autodetect): Ideogram-4 + Qwen-Image-Layered precision-config auto-detect

Build Pipeline [auto-derive] now recognizes two more model families and
applies their per-layer precision config automatically.

Ideogram-4: detect via the unique llm_cond_proj key / Ideogram4Pipeline class.
Its official distribution is an fp8 base the a4w4 recipe is measured on, so any
non-pre-quantized base (fp16/bf16 AND fp8) is eligible; only stamped/nvfp4
exports are skipped. -> QuantFunc/Ideogram-4-Series / ideogram4_a4w4.json.

Qwen-Image-Layered: shares the base QwenImageTransformer2DModel class + block
structure, so it is disambiguated by the unique addition_t_embedding key plus a
shard-independent sibling-config probe (use_additional_t_cond / model_index
QwenImageLayeredPipeline). -> QuantFunc/Qwen-Image-Layered-Series, 50x-above-fp4
(SM120+) / 50x-below-int4 by SM.

Adds tests/ (32 tests; validated against the real models on disk).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>