Add multi-tier slab allocator for ROCm backend

Replace SmallSizePool with a generalized SlabAllocator containing 18
power-of-2 size class pools (8B through 1MB). Each pool pre-allocates
slab pages and sub-allocates via O(1) free-list operations, eliminating
hipExtMallocWithFlags calls for small/medium allocations during decode.

- SizeClassPool: configurable block size, grow-on-demand slab pages
- SlabAllocator: O(1) size-class dispatch via bit ops
- Pre-allocates tiers 0-11 (8B-16KB) at startup (~5.8MB)
- Applies hipMemAdvise/hipMemPrefetchAsync on slab pages
- BufferCache still handles >1MB allocations

Author: Geramy