Per-node concurrency limiter for runtime module compilation.

Compiling is CPU-bound, so allowing more compiles at once than the node has cores only thrashes the scheduler and the (single) code server — it does NOT add resources. This gate caps concurrent compiles to System.schedulers_online() (one per core, the measured sweet spot); extra compiles wait in a FIFO queue and start as slots free. So a cold-start burst of many distinct pages can't saturate the node and stall live requests.

Every compile funnels through run/1, called by the one MishkaCmsCore.Runtime.Handlers.ModuleCompilerHandler worker a module has — covering boot, lazy first-request, and Oban-worker compiles. The callers waiting for that worker hold no slot.

Per-node by design: CPU is a per-node resource, so each node throttles its own compiles with zero cluster coordination. The limit is overridable via config :mishka_cms_core, :compile_gate_limit.

A compile that has to wait for a slot emits [:mishka_cms, :runtime, :compile_gate, :queued] from the gate's process, with how many are waiting as its measurement and the waiting pid as its metadata, once that compile has its place in the queue.