feat: dynamically split PD decode requests

[shihaobai]

Summary

• replace the static LIGHTLLM_PD_SPLIT_MAX_NEW_TOKENS chunking loop with capacity-driven PD segments
• use aggressive EMA-based admission for PD Decode without applying the router-side batch_max_tokens prefill gate
• when an overlapped D request runs out of KV capacity, lower its max_new_tokens to the current accepted output length so the existing post-handle path naturally returns finish_reason=length
• continue the logical request on the same P/D pair so incremental prompts can reuse cache; non-overlap Decode keeps the existing pause/recover behavior

Tests

• 16 passed: PD dynamic split, PD master cached-token handling, and PD selector tests
• Python compile checks for the modified router/backend modules
• git diff --check