Update README for factual correctness (#394)

The README's summary counts and Highlighted Models table had drifted
from the current codebase. Numbers and model listings are now derived
from the actual registry, diffusers class map, task classes, and
component classes.

### Counts line
Corrected the "Supports …" summary, measured from source rather than
hardcoded:

| Metric | Was | Now | Source |
|---|---|---|---|
| Transformers model types | 130 | 290+ | `registry._map` (294) |
| Diffusers component types | 5 | 10 | `_DIFFUSERS_CLASS_MAP` |
| Task types | 14 | 40+ | `ModelTask` subclasses in `tasks.__all__` (41)
|
| Reusable components | 56+ | 100+ | public `nn.Module` classes in
`components` (103) |

`290+` matches the docs generator, which counts the same
`registry._map`.

### Highlighted Models table
- **Text Generation** — added Gemma 4; kept `Qwen 3.6` (same as Qwen
3.5, so supported).
- **Mixture of Experts** — added Qwen3-Next, GLM-4-MoE.
- **Multimodal** — `Gemma 3` → `Gemma 3/4`; removed unregistered
`Phi-3V`; kept `Qwen3.5/3.6-VL`.
- **Speech-to-Text** — added FastConformer-RNNT, FunASR, Qwen3-ASR,
SenseVoice.
- **Diffusion** — added HunyuanDiT, CogVideoX.

### Unchanged
Python API signatures (`build`, `CausalLMTask`), CLI syntax, example
paths, docs links, and skill references were verified against the code
and left as-is.

---------

Co-authored-by: copilot-swe-agent[bot] <198982749+Copilot@users.noreply.github.com>

Author: Copilot

README.md

@@ -25,19 +25,19 @@ multi-component export for pipelines.
 
 | Category | Examples |
 |---|---|
-| **Text Generation** | Llama 2/3/4, Mistral, Qwen 2/2.5/3/3.5/3.6, Phi-3/3.5, Gemma 1/2/3, Granite, GPT-2, OPT, OLMo, SmolLM3, and many more |
-| **Mixture of Experts** | PhiMoE, GPTOSS, Mixtral, OLMoE, DeepSeek-V2/V3, Qwen2-MoE, Qwen3-MoE, Arctic, DBRX, Jamba |
-| **Multimodal** | Gemma 3, Phi-3V, Phi-4MM (vision + audio + LoRA), LLaVA, InternVL2, Qwen2.5-VL, Qwen3-VL, Qwen3.5/3.6-VL, Pixtral |
+| **Text Generation** | Llama 2/3/4, Mistral, Qwen 2/2.5/3/3.5/3.6, Phi-3/3.5, Gemma 1/2/3/4, Granite, GPT-2, OPT, OLMo, SmolLM3, and many more |
+| **Mixture of Experts** | PhiMoE, GPTOSS, Mixtral, OLMoE, DeepSeek-V2/V3, Qwen2-MoE, Qwen3-MoE, Qwen3-Next, GLM-4-MoE, Arctic, DBRX, Jamba |
+| **Multimodal** | Gemma 3/4, Phi-4MM (vision + audio + LoRA), LLaVA, InternVL2, Qwen2.5-VL, Qwen3-VL, Qwen3.5/3.6-VL, Pixtral |
 | **Encoder-only** | BERT, RoBERTa, ALBERT, DeBERTa, DistilBERT, ELECTRA, XLNet |
 | **Encoder-Decoder** | BART, T5/mT5, Marian, M2M-100, Pegasus, BigBird-Pegasus |
-| **Speech-to-Text** | Whisper |
+| **Speech-to-Text** | Whisper, FastConformer-RNNT, FunASR, Qwen3-ASR, SenseVoice |
 | **Audio** | Wav2Vec2, HuBERT, WavLM, SpeechT5 |
 | **Vision** | ViT, BEiT, DeiT, DINOv2, Swin, CLIP, SigLIP |
-| **Diffusion** | Stable Diffusion (UNet + VAE + ControlNet), Flux, SD3, DiT, QwenImage |
+| **Diffusion** | Stable Diffusion (UNet + VAE + ControlNet), Flux, SD3, DiT, QwenImage, HunyuanDiT, CogVideoX |
 | **Adapters** | T2I-Adapter, IP-Adapter |
 
-Supports **130 Transformers model types** and **5 Diffusers component types**
-across **14 task types** and **56+ reusable components**.
+Supports **290+ Transformers model types** and **10 Diffusers component types**
+across **40+ task types** and **100+ reusable components**.
 
 See the [model documentation](https://onnxruntime.github.io/mobius/models/index.html) for the complete list.