[LLM:Bugfix] Fix embedding QNN export paths

[ZedingZhang]

Summary

• detect embedding models in the QNN export wrapper and generate sentence_embeddings IO without logits_index
• use embedding.mnn / embedding.mnn.weight for embedding model QNN export and config_qnn.json
• restore embedding exporter output prefix and make generateLlmIO select embedding.mnn with safer qwen3 detection

Fixes #4481

Tests

• python3 -m py_compile transformers/llm/export/npu/generate_llm_qnn.py transformers/llm/export/llmexport.py
• inline Python checks for embedding and normal LLM input.json generation
• cmake -S . -B build -DMNN_BUILD_LLM=ON -DMNN_LLM_BUILD_DEMO=ON -DMNN_BUILD_TEST=OFF -DMNN_BUILD_BENCHMARK=OFF -DMNN_BUILD_DEMO=OFF -DMNN_BUILD_CONVERTER=OFF -DMNN_BUILD_TRAIN=OFF -DMNN_BUILD_QUANTOOLS=OFF
• cmake --build build --target generateLlmIO -j4

[CLAassistant]

All committers have signed the CLA.

[ZedingZhang]

Closing for now to run local Qwen3-Embedding-0.6B export validation before reopening.

[ZedingZhang]

Local validation completed for this PR: Qwen3-Embedding-0.6B QNN export passed in a linux/amd64 Docker container with QAIRT/QNN SDK 2.47.0.260601. Verified outputs include qnn/embedding.mnn, embedding.mnn.weight, and 29 graph*.bin context binaries. config_qnn.json points to qnn/embedding.mnn and embedding.mnn.weight with chunk_limits [128,1].