Fix eviction queue dead_nodes counter underflow causing excessive purge iterations

[krleonid]

Summary

• Change total_dead_nodes from atomic<idx_t> (unsigned) to atomic<int64_t> (signed) to prevent underflow
• Fix Purge() loop early-out condition to treat negative dead_nodes as zero
• Remove unnecessary UnsafeNumericCast in duckdb_eviction_queues output

Problem

The total_dead_nodes counter underflows because DecrementDeadNodes() is called for nodes that were never counted by IncrementDeadNodes() — specifically, nodes that die (weak_ptr expires or block is pinned) without ever being superseded by a newer version.

When the unsigned counter wraps to a massive value, the Purge() loop's condition 2.2 never triggers:

idx_t approx_dead_nodes = total_dead_nodes;  // underflowed to huge unsigned value
approx_dead_nodes = approx_dead_nodes > approx_q_size ? approx_q_size : approx_dead_nodes;
// ^ clamped to approx_q_size

idx_t approx_alive_nodes = approx_q_size - approx_dead_nodes;
// ^ = approx_q_size - approx_q_size = 0

if (approx_alive_nodes * (ALIVE_NODE_MULTIPLIER - 1) > approx_dead_nodes)
// ^ 0 * 3 > approx_dead_nodes → always false! Loop never exits via condition 2.2

This causes the purge loop to run all max_purges iterations (up to approx_q_size / 8192) unnecessarily, holding purge_lock the entire time.

Observed output from duckdb_eviction_queues

{
    "data": [
        {"approximate_size": 5098963, "dead_nodes": 3873616, "queue_index": 0, "queue_type": "BLOCK_AND_EXTERNAL_FILE", "total_insertions": 605198908},
        {"approximate_size": 224455537, "dead_nodes": 163476417, "queue_index": 1, "queue_type": "MANAGED_BUFFER", "total_insertions": 1977873276},
        {"approximate_size": 29715, "dead_nodes": -107922767, "queue_index": 2, "queue_type": "MANAGED_BUFFER", "total_insertions": 108300237},
        {"approximate_size": 32337, "dead_nodes": -134953305, "queue_index": 3, "queue_type": "MANAGED_BUFFER", "total_insertions": 135275557},
        {"approximate_size": 30044, "dead_nodes": -73130179, "queue_index": 4, "queue_type": "MANAGED_BUFFER", "total_insertions": 73253127},
        {"approximate_size": 28361, "dead_nodes": -32992668, "queue_index": 5, "queue_type": "MANAGED_BUFFER", "total_insertions": 33069332},
        {"approximate_size": 33218, "dead_nodes": -171896118, "queue_index": 6, "queue_type": "MANAGED_BUFFER", "total_insertions": 172673075},
        {"approximate_size": 0, "dead_nodes": 0, "queue_index": 7, "queue_type": "TINY_BUFFER", "total_insertions": 0}
    ]
}

Queues 2-6 show negative dead_nodes — these are the underflowed counters causing excessive purge iterations on every trigger.

Test plan

• All eviction queue tests pass (*eviction* — 10 test cases, 4095 assertions)
• Buffer storage tests pass (test/sql/storage/buffer* — 4 test cases)
• Parallelism tests pass (test/sql/parallelism/* — 68 test cases, 412018 assertions)
• CI full test suite