Add opt-in drainProducerRate to throttle producers during the backlog drain

[oleiman]

buildAndDrainBacklog() only pauses/resumes consumers — producers keep publishing
at producerRate through the drain. That makes it impossible to measure max
consume throughput on a cold backlog (e.g. one aged onto tiered storage) without
producers competing.

This adds an opt-in Workload.drainProducerRate (int, default -1 = unset).
When ≥ 0, buildAndDrainBacklog() calls the existing worker.adjustPublishRate()
just before resuming consumers:

• unset / -1 → producers run through the drain (unchanged)
• N → throttled to N msg/s; 0 → ~1 msg/s via the existing <1 floor (≈ stopped)

It reuses the internal lever findMaximumSustainableRate already uses (fans out to
distributed workers) — no new Worker method or endpoint. Default -1 skips the
call, so existing workloads are byte-identical.

Tested: WorkloadTest covers the default + deserialization; a backlog run with
drainProducerRate: 0 drops producers to ~1 msg/s at "Start draining backlog".

Requires https://github.com/redpanda-data/vtools/pull/4353 to use in benchrunner

[ballard26]

LGTM