fix(server): port eval soft/hard think-close reply budgets

[paride-cmd]

Summary

• Ports `ds4-eval`'s soft/hard reply-budget + forced `` controller into `ds4-server` decode.
• Adds `--soft-limit-reply-budget` (default 1024), `--hard-limit-reply-budget` (default 512), and `--soft-limit-think-close-rank` (default 3).
• Soft-close accepts the model's own `` when it is already in the top-N logits; hard-close forces the close marker to leave answer room.
• Enabled only when request `max_tokens > hard_limit` (same shape as eval's hard < tokens guard). Forced close skips MTP speculation for that step.

Fixes #895

Test plan

• `make ds4_test && ./ds4_test --server`
• `./ds4-server --help thinking` shows the new flags
• Live HTTP: saturating thinking prompt with small `max_tokens` should emit `` and an answer instead of hanging until client timeout

AI usage disclosure: YES — assisted by Cursor agent; logic mirrored from `ds4_eval.c` and covered by new parse/forced-token unit tests.

Made with Cursor