perf: use power-of-two random choices for channel selection to avoid thundering herd

[rahul2393]

Summary

Shared background executor

• Replace per-pool ScheduledExecutorService with a single static ScheduledThreadPoolExecutor (SHARED_BACKGROUND_SERVICE) shared across all GcpManagedChannel instances in the process.
• Each pool now holds ScheduledFuture<?> handles (cleanupTask, scaleDownTask, logMetricsTask) and cancels them on shutdown()/shutdownNow() instead of terminating a dedicated thread pool.
• removeOnCancelPolicy(true) ensures cancelled tasks are immediately purged from the work queue.
• Thread count: max(2, min(4, availableProcessors / 2)).

Power-of-two channel selection (thundering herd fix)

• Problem: pickLeastBusyChannel used a deterministic linear scan that always picked the first channel on tie. Under burst traffic — especially at startup or after idle periods when stream counts are equal — all concurrent callers see the same counts before any are incremented (TOCTOU race), causing all requests to pile onto channel 0. In steady-state with smooth traffic the existing scan distributes fine; the issue is acute during bursts when channels have equalized.
• Fix: Default channel selection now uses the "power of two random choices" algorithm — sample two random channels and pick the less busy one. When stream counts are tied (common in low traffic), prefer the channel with the most recent lastResponseNanos to preserve connection warmth.
• Backward compatibility: The previous deterministic behavior is available as ChannelPickStrategy.LINEAR_SCAN via GcpChannelPoolOptions.setChannelPickStrategy(). The new default is POWER_OF_TWO.

Design decisions

• Non-fallback path uses the configured strategy (POWER_OF_TWO or LINEAR_SCAN).
• Fallback-enabled path always uses linear scan — it must filter channels by fallbackMap and DEFAULT_MAX_STREAM, which requires a full scan.
• Scale-up with POWER_OF_TWO uses getMaxActiveStreams() (the global maximum across all channels) — if ANY channel hits the watermark, we scale up before other channels follow. This is more aggressive than using the global minimum or sampled minimum, but maxSize guards against over-provisioning. The global min would delay scale-up; the sampled min would be noisy.
• Scale-down is unaffected — it already uses a pool-wide aggregate (totalActiveStreams), which is distribution-agnostic.
• Warm tie-breaking (lastResponseNanos) naturally concentrates traffic on active channels under low load without any threshold or mode-switch logic.

[rahul2393]

cc: @kinsaurralde

[olavloite]

For a follow-up PR: This field (and probably most of the other fields here) can be made final.

[rahul2393]

Agreed, will address in a follow-up. Most of these fields are set once in initOptions() during construction and never mutated after.