BufferPool: add size-class bucketing and eviction

[m96-chan]

Current issues with gpu/buffer-pool.ts:

1. Linear scan (O(n)): acquire() walks the entire pool array for matching usage
2. No best-fit: size >= aligned check means a 256-byte request can grab a 1GB buffer
3. No eviction: Pool only grows, never shrinks — no cleanup of unused buffers

Proposed improvements

• Size-class bucketing (power-of-2) for O(1) acquire
• Best-fit or smallest-fit selection to prevent GPU memory waste
• High-watermark based eviction to reclaim unused buffers