fix: smooth txgen send dispatch

[mattsse]

Summary

This changes the txgen sender so queued workload transactions keep refilling available RPC permits instead of stopping dispatch when all permits are temporarily occupied.

Why

High-rate runs could collapse into visible submission bursts because the sender stopped pumping at the concurrency ceiling and the rate limiter accumulated a bounded catch-up burst. At 50k TPS, the old 10ms burst budget could dispatch roughly 500 transactions at once, producing a sawtooth inflight graph.

The first smoothing attempt used one fixed timer interval per transaction, but at high TPS that required sub-millisecond wakes, e.g. 20us at 50k TPS. Runtime timer granularity makes that cap out around millisecond-scale dispatch instead of sustaining the target rate.

Impact

The sender now waits for a freed permit when ready transactions are queued, drains completions, reselects a ready transaction, and dispatches immediately. Rate limiting uses 1ms batches capped to one tick of credit, so 50k TPS releases about 50 sends per tick instead of either a 500-send burst or one 20us sleep per send.

Root Cause

The previous token-bucket limiter intentionally allowed a 10ms burst, and the pump loop returned when permits were exhausted instead of remaining active to refill newly freed permits.