Add real Ideogram 4 unconditional transformer support

[bghira]

What

Ideogram 4 performs CFG through a second, image-only unconditional transformer with zeroed text conditioning. Until now SimpleTuner approximated the unconditional branch by running negative prompts through the conditional transformer, which produces structurally different guidance — particularly visible as noisy targets when AnyFlow distillation fuses guidance at higher scales.

• --ideogram_load_unconditional_transformer loads the frozen unconditional transformer through the same indexed state-dict machinery as the conditional path, respecting --ideogram_fp8_base_upcast, and wires it into both pipeline constructions so validation runs the real asymmetric-CFG branch.
• --ideogram_uncond_ramtorch builds the unconditional transformer on CPU and streams its layers to the accelerator per forward pass via the shared _apply_ramtorch_layers helper (new force flag bypasses the global --ramtorch gate), for fitting both transformers on smaller cards. VRAM cost of the resident option: ~10 GiB fp8, ~18 GiB upcast.
• AnyFlow's _unconditional_batch marks batches with is_unconditional_pass, and Ideogram4.model_predict dispatches marked batches to the unconditional transformer with zeroed llm conditioning, so fused-guidance distillation trains against the model's true unconditional output.
• Quickstart Validation section documents both options (en, es, ja, zh, hi, pt-BR).

Testing

.venv/bin/python -m unittest tests.test_ideogram_unconditional tests.test_ideogram_fp8_linear tests.helpers.distillation.test_anyflow_distiller -v -f — 75 tests pass (1 environment-dependent skip). Coverage includes flag-off/flag-on loading, fp8 upcast handling, the ramtorch offload path, model_predict dispatch geometry, and the unconditional batch marker.