orchestrator: duration fields use three inconsistent serde formats

Problem

Duration/time fields across the orchestrator config files use three different serde encodings, so a user editing the config has to remember a different syntax per field. All three appear in files a user routinely touches (settings.yml, client-parameters.yml, node-parameters.yml):

Style	Encoding	Example field(s)	YAML syntax
1. DurationSeconds (serde_with)	integer seconds	Settings::benchmark_duration, scrape_interval, ssh_timeout, health_check_timeout (crates/orchestrator/src/settings.rs)	benchmark_duration: 300
2. humantime_serde	human string	LoadGeneratorConfig::initial_delay (crates/replica/src/config.rs)	initial_delay: 30s
3. default Duration serde	{secs, nanos} struct	DagParameters::round_timeout (crates/dag/src/config.rs)	round_timeout: {secs: 1, nanos: 0}

So a 1-second value is written three different ways depending on the field: 1 (as part of seconds), 1s, or {secs: 1, nanos: 0}. This is confusing and error-prone (e.g. easy to write benchmark_duration: 30s and have it fail, or expect round_timeout: 1).

Suggested fix

Standardize on humantime_serde everywhere a Duration is configured — it is the most readable and unambiguous (300s, 30s, 1s, 500ms) and round-trips cleanly. Concretely:

• Replace the #[serde_as(as = "DurationSeconds")] fields in Settings with #[serde(with = "humantime_serde")].
• Replace the default Duration serde on DagParameters::round_timeout (and any other struct-form durations) with humantime_serde (use humantime_serde::option for the Option<Duration>).
• Update the asset templates and docs accordingly.

Alternatively, pick any single style — the key is consistency.

Found while setting up multi-region benchmark configs.