Investigate potential head-of-line blocking in ShapeLogCollector for shapes with subqueries

[alco]

Feedback from a user indicates that a shape query with a few nesting levels takes orders of magnitude longer to return than how long it takes to run the query against Postgres (5-8 seconds vs 20ms).

The Electric instance is running at 1-3% CPU utilization on 4 vCPUs with 3.8% memory utilization. The write pipeline processes each transaction in 2.6ms average (8.2ms max). The SLC publish phase takes 1.8ms average (6.3ms max).

The bottleneck is not resource exhaustion — it's architectural. The SLC processes all transactions through a single GenServer with :infinity timeout, creating head-of-line blocking proportional to active shape count and dependency layer depth.