feat: add PixArt model family support

[losewayy]

Summary

Closes #1965.

Add support for the PixArt-Sigma transformer (DiT architecture with adaLN-single modulation):

• New runner src/model/diffusion/pixart.hpp: patch embed via strided conv, 2D sincos position embedding with multi-resolution coordinate scaling, 28× adaLN-single blocks (self-attn + T5 cross-attn + ffn), learned-variance output (8ch → 4ch slice). Config is detected from weight names/shapes rather than hardcoded.
• Detection: PixArt shares the adaln_single.emb.timestep_embedder signature with LTX-AV; disambiguated by pos_embed.proj which is PixArt-only.
• Sigma schedule: PixArt trains with linear beta (0.0001 → 0.02), added calculate_alphas_cumprod_linear_beta used when version is PixArt.
• VAE: PixArt uses the standard SD1.x autoencoder (z=4, scale_factor=0.18215) — excluded from the DiT VAE branch and added to the SD1/SD2 scale-factor branch.
• T5 tokenizer fix: empty input no longer emits a stray ▁ piece; uncond conditioning now produces <eos> + padding, matching Hugging Face. This affects all T5-based models' negative-prompt path — happy to split into a separate PR if preferred.
• Quantization: scale_shift_table is excluded from conversion (element-wise views are invalid on quantized block layouts, and modulation params shouldn't lose precision anyway — same latent issue exists in LTXV). A defensive ggml_cast dequantizes the table if a pre-quantized file is loaded.

Tested with PixArt-Sigma-XL-2-1024-MS (standalone diffusers-format transformer + T5-XXL + SD VAE):

sd-cli --diffusion-model transformer.safetensors --t5xxl t5xxl.safetensors --vae vae.safetensors -W 1024 -H 1024 ...

Test plan

• 1024×1024 generation matches diffusers PixArtSigmaPipeline reference (same composition, photo quality)
• Single-forward numerical parity vs diffusers: max diff ≤ 0.005
• T5 embeddings (cond + uncond) parity vs HF: max diff ≤ 0.008
• q8_0 conversion + inference; legacy fully-quantized gguf also runs via defensive dequant
• --diffusion-fa path OK
• SD1.5 regression: end-to-end generation unchanged
• 512-MS checkpoint (arch identical, needs --model-args for pos-embed base size — documented)
• Other backends/OS: covered by CI

[leejet]

Thank you for your contribution.