Add text-generation-v2 demo with dropdown model selector and non-GQA,GPU-resident KV support

[arisha07]

No description provided.

[arisha07]

Adds a new demos/text-generation-v2/ demo alongside the existing text-generation demo. It introduces a dropdown model selector with an explicit "Load Model" button (no auto-load), and extends LLM inference to support both GQA (fused) and non-GQA (standard attention) models with GPU-resident KV cache.

What's new

1. Dropdown model selector : models are defined in a MODELS config object; adding a new model requires only a config entry, no HTML changes
2. Non-GQA model support : Models generated from Optimum and NNCF flow alongside GQA fused models in the same demo
3. GPU-resident KV cache for non-GQA : a standalone WebNN slice graph dispatches present[1:] → past each step entirely on GPU (no GPU→CPU→GPU roundtrip); only logits (~2.4MB) are read back to CPU per step vs ~134MB with CPU KV
4. freeDimensionOverrides for non-GQA : all dims set statically at session creation, eliminating dynamic-shape dispatch failures

Non-GQA onnx models are available - https://gfx-assets.fm.intel.com/artifactory/gfx-ort-ovep-assets/genai/models/onnx/
For testing use models which are less than 2GB in size. Please update the model path accordingly in main.js

URL : http://localhost:8080/demos/text-generation-v2/?devicetype=gpu this will load ort wasm lib from local dist present in /demos/text-generation-v2/dist.