Title: feat: add bin-packing scheduling strategy for sandbox placement

Summary

Add an optional bin-packing scheduling strategy alongside the existing spread (Best-of-K) strategy. This allows operators to choose between two placement modes via a feature flag:

• Spread (default): Distribute sandboxes evenly across all nodes — best for high availability and fault isolation
• Bin-pack: Fill one node before moving to the next — best for cost optimization and resource consolidation

Motivation

The current Best-of-K placement algorithm always prefers nodes with lower utilization (spread strategy). This is ideal for large-scale deployments where fault isolation matters, but suboptimal for:

1. Cost-sensitive self-hosted deployments: When running on auto-scaling infrastructure, bin-packing keeps fewer nodes active, allowing idle nodes to be scaled down and reducing cloud costs
2. Small clusters (2-5 nodes): Spreading 10 sandboxes across 5 nodes means all 5 nodes must stay running. Bin-packing would concentrate them on 1-2 nodes, freeing the rest
3. Dev/staging environments: Where availability is less critical than cost

Proposed Design

• Add a feature flag sandbox-placement-bin-pack (default: false)
• When enabled, reverse the scoring logic in BestOfK.Score(): prefer nodes with higher utilization (more allocated resources)
• In bin-pack mode, evaluate all eligible nodes instead of sampling K random ones, ensuring the most-loaded node is always selected
• No API changes required — purely a server-side scheduling optimization

Behavior Comparison

Scenario (3 nodes, 6 sbxs) Spread Bin-pack
Node A 2 sandboxes 4 sandboxes
Node B 2 sandboxes 2 sandboxes
Node C 2 sandboxes 0 (can scale down)

/cc @jakubno

Locate the existing BestOfK.Score() implementation and the sandbox placement feature-flag configuration. Read how eligible nodes are sampled and scored, then run the scheduler tests that cover placement behavior. Done means the default spread strategy is unchanged, bin-pack evaluates all eligible nodes and prefers higher utilization, and the feature flag remains disabled by default.