THROTTLED_AT_CONSENSUS is issued with throttles set at 50K TPS

### Description

When running the load generator at full speed, throttles are artificially set to 50K to effectively disable them. The network occasionally issues `THROTTLED_AT_CONSENSUS` even when the actual TPS is 10-15K. Transactions are not throttled at frontend.
This behavior affects measured performance and may create poor user experience.

### Steps to reproduce

Run `NftTransferLoadTest` in a Latitude network at large scale (100M+ accounts/NFTs).

### Additional context

Backend throttling uses consensus time which is the same for all transactions in a round. The size of a round in terms of events/transactions may fluctuate a lot. A general algorithm that works well for frontend throttling may not be adequate for such "bursts".

### Hedera network

other

### Version

v0.57.0-SNAPSHOT

### Operating system

Linux

Start by reproducing the behavior with NftTransferLoadTest in a large-scale Latitude network, then trace backend throttling that uses consensus time and compare it with frontend throttling. Done means the load test no longer reports spurious THROTTLED_AT_CONSENSUS results when actual throughput is 10–15K TPS.

Written by the indexing model from the issue text.