feat: add rec_type and onerec batch input builder for rec framework[4/8].

[DragonFive]

This PR is a follow-up to #557 and continues the generative recommendation work by introducing a RecType-based
orchestration layer and a layered rec batch input builder.

What’s included

• RecType plumbing (kNone/kOneRec/kLlmRec) from RequestState -> SequenceParams -> Sequence (default: kNone)
• Rec batch input builder layering:
  • RecBatchInputBuilder (abstract + factory)
  • OneRecBatchInputBuilder (concrete implementation for kOneRec)
  • Batch selects the builder based on RecType
• Scheduler wiring: use create_rec_batches for the rec scheduling path and tag rec requests as RecType::kOneRec
• Sequence behavior: OneRec-specific output is gated by rec_type == kOneRec; other types default to the standard LLM
  text decode path

Out of scope (planned follow-ups)

• Rec worker (runtime) + input/output processors (request ingestion and result emission)
• Concrete kLlmRec implementation (enum value reserved only)
• OneRec model implementation + ATB graph building/integration
• Pure NPU inference path

Roadmap checklist

• Rec tokenizer feat: add generative recommendation tokenizer. #317
• Rec decoder constrain feat: add constrained decoding for generative recommendation #480
• Rec proto and service feat: add rec proto,serivce and utils for rec framework[2/6]. #440
• Rec engine master feat: add rec scheduler master and engine for rec framework[3/8]. #557
• Scheduler feat: add rec scheduler master and engine for rec framework[3/8]. #557
• RecType plumbing + rec batch input builder layering feat: add rec_type and onerec batch input builder for rec framework[4/8]. #565
• Rec worker
• request+sampler
• OneRec model + ATB graph integration
• Pure NPU inference

[DragonFive]

when #557 is merged, i will rebase this branch and push again.