made txs shard handling stricter

[pompon0]

Txs in each shard will be sent sequentially, which is consistent with what autobahn expects.

• I have separated the txs sharding from endpoints - they are separate concepts now
• each endpoint will allow sending up to tasksPerEndpoint transactions in parallel, as before
• endpoint for each shard is sticky for backward compatible behavior
• the number of shards increases by default from len(endpoints) to tasksPerEndpoint * len(endpoints)
• introduced a new config field NumShards which allows configuring the number of shards independently from the number of endpoints.
• since there will be significantly more shards, we cannot afford each of them to have a large queue any more, so the queue capacity is shared now. I have introduced a QueuePool for that purpose - it maintains a collection of queues with a shared capacity, sending/receiving from each queue is gated separately for efficient goroutine management.
• I have updated the utils package by copying over (part of) the latest version of sei-chain/sei-tendermint/libs/utils

[cursor]

PR Summary

High Risk
This refactors the core transaction send pipeline (queueing, sharding defaults, and closed- vs open-loop rate limiting), so throughput, backpressure, and latency behavior can change materially on real load runs.

Overview
Stricter sharding for load generation: txs are routed to N shards (default endpoints × tasksPerEndpoint, overridable via numShards) instead of one queue per endpoint. Each shard drains sequentially through a dedicated goroutine; endpoint assignment stays shardIndex % len(endpoints) while each endpoint still runs up to tasksPerEndpoint parallel RPC senders via the new ethClient.

The old per-endpoint Worker (buffered channel + optional skip-rate-limit) is replaced by QueuePool (one shared backlog cap across all shards), ShardedSender shard loops that limiter.Wait before RPC, and ethClient for HTTP/WS sends, metrics, and inclusion registration.

Open-loop: when the scheduler owns the arrival clock, main passes an unlimited sender limiter so TPS is not double-gated; prewarm always self-paces on the shared limiter. The open-loop scheduler now uses golang.org/x/sync/semaphore instead of a custom helper.

Plumbing: concurrency moves to utils/scope (with utils/service as thin aliases); utils gains RWMutex, OrPanic, and related helpers synced from sei-tendermint; queue-depth metrics aggregate ShardStats from registered ShardedSender instances.

^{Reviewed by Cursor Bugbot for commit 25d018e. Bugbot is set up for automated code reviews on this repo. Configure here.}

[wen-coding]

Since this is SendOrDrop, under load we may drop some txs from receipt tracking. The tx is still submitted, but we'd never observe an on-chain failure for the dropped ones — so TrackReceipts is effectively a sample, not full coverage. Is that correct?

[pompon0]

it is. Also, afaiu in case of autobahn it will not be available at all since receipt store is not ready, right?

[wen-coding]

In Autobahn the receipt store currently works for EVM tx (although there are many performance and safety issues the storage team is fixing), it doesn't work for Cosmos tx at all.

[pompon0]

anyway, after rebase receipt requests are gone entirely in favor of observing the finalized blocks.

[cursor]

Cursor Bugbot has reviewed your changes using default effort and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit ead64e6. Configure here.}