P1: Add LLM burst guard (concurrency caps + jittered queueing)

[bleuropa]

Summary

Add burst guard controls to reduce provider 429 spikes during multi-agent fan-out.

Context

Current behavior uses retry/backoff and token bucket acquisition, but lacks request-concurrency shaping at burst time.

Relevant code:

• lib/loomkin/teams/rate_limiter.ex#L23-L80
• lib/loomkin/llm_retry.ex#L43-L48

Scope

• Introduce per-provider and per-team in-flight request caps.
• Queue with jittered dispatch to avoid synchronized retries.
• Keep existing retry behavior but reduce coordinated burst pressure.

Acceptance Criteria

• Under synthetic multi-agent load, 429 rate decreases compared to baseline.
• Requests are queued and drained predictably instead of burst-failing.
• Telemetry exposes queue depth, wait time, and throttling events.