faster-redis: native Zig connection pool and hot-path buffer reuse

[justrach]

Move pooling from Python wrappers into the Zig extension so free-threaded workloads stop paying Python queue/checkout overhead.

Progress:

• added native pool handle support in zig/src/main.zig
• added pooled execute / pooled pipeline entrypoints
• added per-connection reusable write buffers in Zig pack paths
• fixed RESP array lifetime cleanup while touching the hot path
• Python PooledRedis now uses the native pool instead of queue.LifoQueue
• refined pool selection from blind round-robin to sticky thread-affine slot assignment
• added cheaper Python response fast-path handling for the hottest command families (GET, SET, MSET, common bool-like replies)
• removed the useless recursive Python list walk in _dec
• added a thread-bound pooled client view so worker threads stop routing every operation back through the shared PooledRedis dispatcher

Validation:

• PYTHONPATH=. /Users/rachpradhan/turboAPI/faster-redis/.venv314t/bin/python -m unittest tests.test_client_fallback tests.test_native_shadow tests.test_parity_harness passes
• isolated local Redis benchmark now shows pooled reads and writes at or above TLS on the tested thread counts

Latest benchmark snapshots against isolated Redis on port 6381:

• 16 threads, pool 16: GET 2.80x vs redis-py, SET 1.20x; TLS is 2.59x on GET and 1.43x on SET
• 32 threads, pool 32: GET 2.61x vs redis-py, SET 2.22x; TLS is 2.52x on GET and 2.20x on SET

Takeaway:

• thread-bound pooled client views were the missing piece for read-heavy scaling
• the remaining difference versus TLS is now gone on the tested pool/thread matches
• pooled design now keeps the shared native pool architecture while matching or slightly beating TLS in these benchmark shapes