Speech RL recipe update

[yuekaizhang]

• Update docker image docker pull soar97/verl:app-verl0.4-vllm0.8.5-mcore0.12.2-te2.2
• Changed datasets from aishell3 to emilia_zh
• Fixed WER computation scripts, support compute MER for code-switch cases
• Fixed cosyvoice3 zero_shot_zh input text normalization
• Reran experiment using temperature=1.0 and top_p=1, disable algorithm.norm_adv_by_std_in_grpo to avoid difficulty bias
• Updated reward function and metrics (disable errors caused by tone.)
• Add dapo training recipe (should work with verl-project/verl@75de3de)

Model	Seed-TTS test_zh CER	Cosyvoice3 zero_shot_zh	Comment
SFT (initialized from Qwen2-0.5B-Instruct)	1.70 %	4.26%	See PR #1887
GRPO (this work, trained on AIShell-3)	1.06 %	3.01%	Commit
GRPO (this work, trained on emilia_zh subset, using top_p=1, temperature=1.0)	0.87%	2.63% (1800 steps)
DAPO (this work, trained on emilia_zh subset, using top_p=1, temperature=1.0)	0.83%	2.71% (700 steps)	See run_dapo.sh

[sharonyu-115]

LGTM. Thank you!