Runtime: actor pooling (demand-gated), reuse actor allocations under spawn/destroy churn

[nicolas-maman]

Context

PR #1330 removed the previous actor-pool machinery because it was inert and hazardous rather than an optimization:

• actor_pool_acquire had zero call sites, so no actor was ever pooled; every core still paid for an initialized 64-slot ActorPool.
• The release path cast ActorBase* to PooledActor*. The two structs are not layout-compatible (the pool_index read lands inside the step function pointer), and when the overlaid bytes landed in 0..63 the miss-branch routed NUMA-allocated memory to plain free() (mmap-backed under libnuma, VirtualAlloc-backed on Windows NUMA).
• The pool held fixed-size slots with their own embedded Mailbox, while real actors are variable-size derived structs (sizeof(PingActor) etc.), so the design could not have been wired without a rewrite.
• Surface that lied: Actor Pooling [ON] in the verbose config print, a documented AETHER_ACTOR_POOL_SIZE env var that fed nothing, a never-incremented actors_pooled counter.

The idea itself is legitimate: high-churn workloads that spawn and retire actors in tight loops (skynet-style trees, per-request actors) pay malloc/free per actor. This issue is the place to decide whether that cost is real and, if so, design a correct pool.

Demand gate (do this first)

Per the project rule that perf work needs profiling evidence before implementation:

1. Benchmark spawn/destroy-heavy workloads (skynet in benchmarks/cross-language/, per-request actor dispatch in std.http) and measure the share of time in scheduler_spawn_actor / scheduler_release_actor allocation.
2. If allocation is not a measurable share, close this issue with the numbers.

Design constraints for a real implementation

• Actors are variable-size: a pool must be size-class bucketed or per-actor-type, not fixed-slot.
• NUMA affinity: today spawn allocates on the caller core's node (aether_numa_alloc); reuse must not silently migrate memory across nodes.
• Work stealing migrates actors between cores mid-life, so release can happen on a different core than spawn; a per-core pool needs a same-node return path (or an ownership rule).
• Reuse must re-initialize the mailbox, atomics, alloc_size, and panic/dead flags with the same guarantees a fresh calloc/aether_numa_alloc gives.
• AETHER_PROFILE plumbing: the removed profile constants (micro/small/medium/large actor-pool sizes) can be revived to size the buckets once something consumes them.
• The actors_pooled stats counter returns only when it can actually increment.

Acceptance

• Measured improvement on the churn benchmarks, no regression on the message-passing suite.
• Valgrind + ASan clean under spawn/destroy stress.
• Documented in docs/runtime-optimizations.md with the activation tier stated honestly.