Memory-pressure-driven cache eviction.

Periodically polls a pluggable pressure source (erllama_pressure) and, when the used/total ratio crosses high_watermark, asks the cache to evict slabs until the ratio would drop below low_watermark. The eviction call is erllama_cache:evict_bytes/2 with a target of (high - low) * Total bytes; the cache may free less if no evictable slabs remain.

Tier policy: by default the scheduler evicts only ram and ram_file slabs. Disk-tier slabs are left in place — disk is the cheap tier, and the deployment usually wants to keep as much warm state as possible there. Disk eviction happens via the cache's own per-tier quota or via an explicit erllama_cache:gc/0 call. Override with evict_tiers => all (or a custom list) to include disk in scheduler-driven eviction.

Disabled by default. Enable via the erllama app environment:

{erllama, [
  {scheduler, #{
    enabled         => true,
    pressure_source => system,    %% noop | system | nvidia_smi | {module, M}
    interval_ms     => 5000,
    high_watermark  => 0.85,
    low_watermark   => 0.75,
    min_evict_bytes => 1048576,   %% don't bother with sub-MB targets
    evict_tiers     => [ram, ram_file]
  }}
]}

The scheduler always starts (so it can be enabled at runtime via enable/1), but its timer only fires when enabled = true.