Add Transformers.js v4 compatibility

[rshube]

Renames VL export outputs to match Transformers.js v4 session registry

Transformers.js v4 hardcodes session-to-filename mappings for ImageTextToText models in session_config.js. It expects vision_encoder, embed_tokens, and decoder_model_merged as the ONNX file basenames.

Changes:

• Rename embed_images → vision_encoder and decoder → decoder_model_merged across export, inference, quantization, and tests
• Generate preprocessor_config.json — Transformers.js reads this (flat format), not processor_config.json (nested under image_processor)
• Inline chat_template.jinja into tokenizer_config.json — Transformers.js reads it from there, not from the separate .jinja file. Strips {%- generation -%} / {%- endgeneration -%} blocks which the Transformers.js Jinja parser doesn't support
• Add transformers.js_config.use_external_data_format to config.json — tells Transformers.js which sessions have .onnx_data files to download

get_onnx_file() in session.py falls back to the old names if the new ones aren't found, so existing exports still work.