v2.0.21

buffer_pool: split into read/write pools with proper backpressure

Split unified buffer pool into separate read and write pools to isolate
memory usage and enable better diagnostics:
- Read pool: 128 MB (8192 buffers) for async_read + async_hash
- Write pool: 256 MB (16384 buffers) for async_write

Implement proper backpressure mechanism:
- Use allocate_buffer(exceeded, observer) in async_write
- Return exceeded flag to trigger libtorrent backpressure
- When write pool exceeds high watermark (87.5%), libtorrent stops
  sending more blocks until pool recovers to low watermark (50%)
- This prevents pool exhaustion and maintains stable memory usage

Clean up unified_cache API:
- Remove unused exceeded and observer parameters from insert_write()
- These parameters had no effect (watermark checking was disabled)

Update logging:
- Change buffer pool warnings to debug level to reduce log noise
- Add detailed pool usage statistics every 30 seconds
- Show separate READ/WRITE pool usage percentages

This addresses the write pool exhaustion issue observed at high download
speeds (700+ MB/s) where the pool would reach 100% and trigger sync write
fallbacks. With backpressure, usage stays between 50-87.5%.

Signed-off-by: Date Huang <tjjh89017@hotmail.com>