Inference split: keep host ops between step annotations

[ajassani]

Summary

• Split windows previously kept CPU ops only while they were nested inside each iteration annotation (ts .. ts+dur). Serving stacks launch work after that envelope and before the next step — vLLM compute_logits / lm_head GEMMs are the usual example — so those ops never made it into the extracted trace or GEMM tables.
• Each root's CPU window now runs until the next same-pattern iteration annotation on that thread. GPU kernels still attach by correlation, so they can sit outside the host window. Neighboring steps do not leak in.
• Not every model: this bites dense TP=1 vLLM serving where sample sits outside execute_* (Qwen3-0.6B/8B/14B, Llama-3.1-8B). Other vLLM configs can still keep vocab GEMMs in the split report. SGLang in-graph lm_head is a separate --capture_folder issue.

Fixes #1023

Test plan

• pytest tests/test_split_inference_trace_annotation.py
• Re-split a vLLM serving trace and confirm compute_logits / vocab-sized GEMMs appear in the extracted window and GEMM sheet
• Spot-check an SGLang serving split still isolates step[DECODE|EXTEND] windows

[codecov-commenter]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!