optimize performance: reduce allocations, buffered I/O, zero-copy reads

[carmark]

Summary

Performance optimizations targeting the hot I/O paths.

Read path - eliminate double allocation (HIGH impact)

• Removed redundant make([]byte, tl) before bs.Read() in bsPerformCommand
• Replaced inefficient append loop for zero-fill with direct in-place zeroing
• Result: 1 fewer allocation per read, no slice growth overhead

parseHeader - use command pool (MEDIUM impact)

• Changed &ISCSICommand{} to getCommand() from sync.Pool
• Every SCSI command was allocating a new ISCSICommand on heap
• Result: near-zero allocation for command parsing under steady load

Unmap - shared zero buffer (MEDIUM impact)

• Replaced per-descriptor make([]byte, desc.TL) with shared 1MB zero buffer
• Large unmaps (e.g., 1GB) previously caused 262K allocations, now zero
• Result: O(1) memory usage regardless of unmap size

Network I/O - buffered writer (MEDIUM impact)

• Added 256KB bufio.Writer to iSCSI connections
• Batches small PDU writes into fewer syscalls
• Flush after txHandler completes to ensure data delivery
• Result: fewer write syscalls, better TCP segment utilization

Test plan

• go build ./... passes
• go vet ./... passes
• go test ./... all tests pass
• Local daemon startup + port binding verified