fix(gemma4): the serving streams wait for what setup enqueued

[FeathBow]

Description

Fixes #1029

Both handoffs relied on time. The loader's and the engine's DeviceContexts each own a non-blocking stream with cudarc's event tracking disabled, so nothing orders work across them unless the code does. The loader returned with the expert repack and scale kernels still enqueued on its stream, and the engine created the async prefill lane's stream without synchronizing its own unless CUDA graphs were on. The shard unmap's host cost and the startup work before the first request were what kept the readers behind the writers.

The loader drains before handing over. from_safetensors synchronizes its stream after the unmap, so the host cost still overlaps the kernels and only the residual is paid; the log line and the LoadStats doc now say the expert kernels are drained, and elapsed_ms includes that residual.

The engine synchronizes before the lane exists. The ctx.sync() that sat inside the graph branch runs unconditionally before AsyncPrefillLane::new, so the KV pools, the arena and the warm passes are complete for every stream that follows, graphs or not.

Test Env

• Single GPU (sm_89, x86_64), --release --features gemma4, the pinned Gemma 4 12B checkpoint and the 26B-A4B NVFP4 checkpoint.

Verification

• cargo fmt --check clean; clippy --all-targets -D warnings clean for pegainfer-gemma4; lib tests 49 passed / 23 ignored, unchanged from main.
• Lifecycle gates through the maintainer runner: 5 of 5 (the shared, green and gathered lane lifecycles on the dense profile, the shared and green ones again on the routed profile, which is the checkpoint with experts).
• Same-config serving A/B on the 26B-A4B checkpoint, main vs this branch alternated twice each: four greedy prompts byte-identical across all eight runs (one md5 over text, finish reason and usage), with the default CUDA graphs and with --cuda-graph=false plus PEGAINFER_ASYNC_PREFILL=shared, the latter being the combination that had no ordering before.
• Load time on the routed checkpoint (submission total from the loader's log line, two runs each, warm page cache): graphs on, main 10594 ms [10579, 10609] vs branch 10702 ms [10602, 10801]; graphs off with the lane, main 10645 ms [10641, 10648] vs branch 10701 ms [10695, 10706]. The difference is the residual expert-kernel drain and sits inside the run-to-run spread; the unmap's host cost still covers most of the kernels.