perf: optionally pool stream writer buffers

[jessy-ddog]

Summary

Adds opt-in destination buffer pooling for stream writers. Pooling remains disabled by default.

Purpose

gRPC creates one zstd writer per message. Each writer allocates a destination buffer sized for the input, which creates significant allocation pressure at high throughput.

Changes Made

• Add a concurrency-safe switch for writer buffer pooling.
• Reuse destination buffers through sync.Pool when enabled.
• Return buffers on every Close path.
• Add a focused test and allocation benchmark.

Impact & Blast Radius

• Affected Components: Stream writers created after SetWriterBufferPoolEnabled(true).
• Risk: Low. Existing behavior remains the default.
• Potential Side Effects: Enabled pooling can retain recently used buffers until a garbage collection clears the pool.

How to Test & Measure

• go test ./... -count=1
• go test -race ./... -run ^TestWriterBufferPool -count=1
• go build -tags external_libzstd ./...
• go test -run ^ -bench ^BenchmarkStreamWriterBufferPool -benchmem -count=3

For a 70 KB payload, pooling reduced allocations from about 75,016 B/op to 139 B/op and from 4 allocs/op to 2 allocs/op.

Checklist

• I have self-reviewed my code.