perf: buffer optimizations (sizes, zero-copy, flush)

[Yanstart]

Context

Parent design: docs/plans/2026-03-18-relay-refactoring-design.md - Stage 5/5
Bottlenecks: 64KB buffers below BDP, .to_vec() copies on non-FEC path, flush() after every TCP write.

Changes

5a. Profile-scaled TCP buffers: 128KB terrestrial, 256KB satellite (cap at diminishing returns)
5b. Zero-copy on non-FEC path: use slice reference instead of .to_vec()
5c. Remove flush() on exit side: TCP kernel flushes naturally
Files: bins/quick-entry/src/tunnel.rs, bins/quick-exit/src/tunnel.rs

Acceptance Criteria

• TCP buffer sizes scaled per profile
• No .to_vec() copy when FEC and compression disabled
• No flush() after each TCP write on exit
• 351+ tests pass, CI green
• Proxmox benchmark: +5-10% on top of stages 1-4

Dependencies

Requires #112 (backpressure) merged first. Can be done in parallel with final benchmarks.

Expected Impact

Polish: +5-10% throughput, reduced memory allocations

[Yanstart]

Updated context (March 19)

This is now Phase 3 of the channel-based I/O driver design. Depends on #116 and #117.

After the I/O driver eliminates Mutex contention, buffer optimizations become the final polish:

• Profile-scaled TCP buffers (128-256KB)
• Zero-copy on non-FEC path (eliminate .to_vec())
• Remove flush() on exit side

Expected: +5-10% on top of channel-based improvements.