Add ForcedSequenceLogitProcessor (teacher-forcing custom logit processor)

[DavidBellamy]

What

Adds a general-purpose ForcedSequenceLogitProcessor to python/sglang/srt/sampling/custom_logit_processor.py. Per request, pass custom_params={"forced_output_ids": [int, ...]}; at decode step t it forces the next token to forced_output_ids[t] by masking all other logits, then samples normally past the end of the list.

Why

Lets two server configurations run a byte-identical decode workload so the only variable is the configuration under test (a caching flag, a kernel, a quantization mode), and supports reproducible / replayed decoding. It's a natural sibling of the existing ThinkingBudgetLogitProcessor, DisallowedTokensLogitsProcessor, and DeepseekOCRNoRepeatNGramLogitProcessor in the same file.

Usage

• Launch the server with --enable-custom-logit-processor.
• Per request (native /generate or /v1/chat/completions):
  • custom_logit_processor = ForcedSequenceLogitProcessor.to_str()
  • sampling_params.custom_params = {"forced_output_ids": [...]}
  • max_new_tokens = len(forced_output_ids)

Standalone; no dependency on other in-flight changes.