[AsyncGRPO] aiohttp limits to 100 reqs when max_inflight_tasks > 100

[ggcr]

By default max_inflight_tasks = max_staleness * per_device_train_batch_size * gradient_accumulation_steps * num_processes

AsyncRolloutWorker aiohttp TCP connector limits it to 100 by default which results in vLLM having 100 reqs running even in cases where max_inflight_tasks > 100

# My max_inflight_tasks = 256 and yet I only have 100 vLLM reqs

Engine 000: Avg prompt throughput: 255.0 tokens/s, Avg generation throughput: 4757.7 tokens/s
Running: 100 reqs, Waiting: 0 reqs, GPU KV cache usage: 12.7%, Prefix cache hit rate: 0.0%
         ^^^^^^^^           ^^^^^^

I wonder if this behavior is intended for stability. Setting that limit to max(100, max_inflight_tasks) so that vLLM handles all max_inflight_tasks? User can adjust --max-num-seqs from vLLM to control Running requests from there...

I can move to discussion if needed, thanks in advance