Avoid redundant bulk remote queue notifications

[fallintoplace]

Summary

• Check the result of push_front() in the static thread pool bulk enqueue path.
• Notify a worker only when its submitter queue changes from empty to non-empty.
• This is a follow-up to Avoid redundant single-task remote queue notifications #2178. The single-task paths were optimized there, while the bulk paths were left unchanged.

Benchmark

Focused sender-level schedule | bulk benchmark on Apple arm64 with AppleClang 21, -O3, 100,000 bulk operations per producer, and the median of 5 runs after 1 warmup. Each cell is end-to-end throughput in Mops/s: main -> patch (change).

Workers / agents	1 producer	2 producers	4 producers	8 producers
2 / 2	5.84 -> 6.59 (+13%)	5.78 -> 5.76 (0%)	4.64 -> 5.36 (+15%)	4.09 -> 4.47 (+9%)
4 / 4	3.39 -> 3.59 (+6%)	2.98 -> 3.41 (+15%)	2.77 -> 3.04 (+10%)	2.50 -> 2.76 (+10%)
8 / 8	1.84 -> 2.22 (+21%)	1.81 -> 1.99 (+10%)	1.67 -> 2.05 (+22%)	1.65 -> 1.85 (+12%)

I also checked smaller bulk operations with one agent per submission:

Workers / agents	2 producers	4 producers	8 producers
2 / 1	5.59 -> 6.56 (+17%)	5.71 -> 5.76 (+1%)	4.46 -> 4.82 (+8%)
4 / 1	4.79 -> 5.51 (+15%)	4.45 -> 4.92 (+10%)	4.33 -> 4.65 (+7%)
8 / 1	4.98 -> 5.20 (+4%)	3.98 -> 4.08 (+3%)	3.36 -> 3.34 (-1%)

The gains are largest when several producers share worker queues.

Testing

• git diff --check
• Static thread pool tests: 9 test cases, 18 assertions
• Full test.exec: 336 test cases, 3,343 assertions

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[ericniebler]

/ok to test e1bd7c4