Feature/gradio UI , Added Gradio Web UI with Dual-Speaker Support, Examples, and Stability Fixes

[NemesisGuy]

Add Gradio Web UI with Dual-Speaker Support, Examples, and Stability Fixes

This PR introduces an interactive Gradio web UI for SoulX-Podcast and includes several supporting improvements for a smoother local inference experience and dual-speaker podcast generation.

Key changes

Gradio UI

• gradio_app.py: Main Gradio app supporting Single and Dual Speaker modes.
• webui.py: Optional helper for WebUI/demo purposes.
• Supports both bracketed tags ([S1] / [S2]) and prefix tags (S1: / S2:).
• Silent-audio fallback if reference audio is missing.
• Sample token resolution via JSON tokens (sample1 / sample2).

Examples & Smoke Tests

• Updated README.md with usage instructions and sample JSONs.
• example/gradio/smoke_test.py: Lightweight input validation and optional full inference test.
• Added example dialog JSONs and sample audio under example/.

Engine & Infrastructure

• llm_engine.py: Clean, merged HF/vLLM wrapper. Optional vLLM imports are guarded; HF is the default.
• In-process model caching and safe fallback to avoid duplicate GPU allocations.

Docs & Metadata

• CHANGELOG.md updated with unreleased feature section.
• README.md now documents the Gradio UI and Dual Speaker workflow.

Merge / Housekeeping

• Resolved rebase conflicts, consolidated .gitignore and README.md.
• Ensured all Python modules compile.

Why this change

• Improves developer experience with a local browser UI for experimentation.
• Adds robust dual-speaker handling (tokenized sample references, dialect prompts, upload precedence rules).
• Introduces a small smoke-test harness to validate inputs quickly.

Testing performed

• Syntax checks via python -m py_compile for edited modules.
• Completed rebase and resolved merge conflicts.
• Local testing of Gradio app and smoke test recommended.

Manual testing completed:

• Smoke test passed
• Gradio UI started successfully and model loaded

Usage notes

• Run light smoke test (no model load):

  python example/gradio/smoke_test.py

• Run full smoke test (loads model):

  python example/gradio/smoke_test.py --run --model-path pretrained_models/SoulX-Podcast-1.7B

• Start UI:

  python gradio_app.py

  Open http://localhost:7860
• vLLM is optional; HF engine is used if vLLM is not installed.
• Consider a quick README tidy-up before publishing to remove duplicate sections from merge resolution.