[spec v2]Fix torch gc of future indices

[hnyls2002]

How to reproduce this on small models (llama3-8b on H200)

diff --git a/python/sglang/srt/managers/scheduler.py b/python/sglang/srt/managers/scheduler.py
index 3435fcaef..526bf04df 100644
--- a/python/sglang/srt/managers/scheduler.py
+++ b/python/sglang/srt/managers/scheduler.py
@@ -2332,6 +2332,7 @@ class Scheduler(
 
                 with self.forward_stream_ctx:
                     self.forward_stream.wait_stream(self.default_stream)
+                    torch.cuda._sleep(1_000_000_000)
                     self.future_map.resolve_future(model_worker_batch)
                     with self.record_forward_metrics(batch):
                         batch_result = self.model_worker.forward_batch_generation(

fix #18744
close #18803

I seriously thought about @nvcastet's and @trevor-m's analysis of the data races, and I concluded that there are actually no data races in the traditional sense. Even though prepare_for_decode and _draft_extend_for_decode can access the shared buffer req_to_token at the same time, there are no conflicts between these two phases — even if there were, they wouldn't cause out-of-bound errors or IMA.
So I started thinking about whether some tensors could be garbage-collected because they were not recorded across streams. Some tensors are created on the scheduler (default) stream but used on the forward_stream, and their reference counts drop to zero during forwarding.
I originally thought the problematic tensor would be a direct field of ScheduleBatch, so I did a full clone of all GPU tensors in ModelWorkerBatch as an ablation. Even after that, the IMA still occurred. Trevor gave me a very useful hint: the indices inside FutureMap had bad values that didn't make sense. That pointed me to the root cause — future_indices.indices is allocated on the default stream, used on the forward_stream, and its Python references are dropped (when model_worker_batch.spec_info and batch.spec_info are replaced) before the GPU finishes reading it. The fix is a record_stream call on the indices tensor.

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

[hnyls2002]

/tag-and-rerun-ci

[trevor-m]

Confirmed that this fixed the issue on our wideep disagg repro