Performance improvements: tunes sysctl configuration and adds many tunnels per interface

[luiscape]

vprox tunnels are notably slow. In an experiment using iperf3, they produced the following results:

32 streams

===================================================================================================================
  Tunnel Throughput (vprox - before changes)
===================================================================================================================
-------------------------------------------------------------------------------------------------------------------
  [tunnel] TCP upload x32                         921.1 Mbps      1098.4 MB  403968 retrans  cpu tx/rx 2%/35%
  [tunnel] TCP download x32                        1.70 Gbps      2023.0 MB       0 retrans  cpu tx/rx 20%/15%
-------------------------------------------------------------------------------------------------------------------

=========================================================================================================
  Baseline Throughput (direct)
=========================================================================================================
---------------------------------------------------------------------------------------------------------
  [baseline] TCP upload x32                4.91 Gbps      5849.0 MB     204 retrans  cpu tx/rx 8%/98%
  [baseline] TCP download x32             11.95 Gbps     14244.9 MB       0 retrans  cpu tx/rx 111%/54%
---------------------------------------------------------------------------------------------------------

vprox is ~4x slower for uploads and ~7x slower for downloads.

Some overhead from encryption is expected when using WireGuard, particularly if the CPU is saturated with encryption or decryption operations. However, a difference this large is unacceptable.

With the changes in this PR, a vprox interface achieves the following throughput in the same experiment:

=========================================================================================================
  Tunnel Throughput (vprox - after changes)
=========================================================================================================
---------------------------------------------------------------------------------------------------------
  [tunnel] TCP upload x32                  3.94 Gbps      4694.6 MB  138560 retrans  cpu tx/rx 7%/92%
  [tunnel] TCP download x32                4.30 Gbps      5130.2 MB       0 retrans  cpu tx/rx 103%/25%
---------------------------------------------------------------------------------------------------------

Uploads reach 3.94 Gbps (80% of baseline) and downloads reach 4.30 Gbps (36% of baseline). This represents a 4x improvement in upload speed and a 2.5x improvement in download speed over the previous version. Both are currently bottlenecked by server CPU utilization — adding more resources will increase throughput further.

This is achieved through two improvements:

• sysctl performance tuning that configures the Linux kernel to use RX/TX queues
• the use of multiple WireGuard tunnels nested under a dummy interface, which increases throughput linearly up to the limits of the NIC and CPU on both the server and client

The client now accepts a --tunnels N parameter, where N is the number of tunnels to start. The client creates a dummy interface that routes traffic across all tunnels in round-robin fashion.

Client                                              Server
┌──────────┐  policy routing    
│  vprox0  │  (dummy, user-     
│          │   facing)          
│          │                    
│          │  ip rule: to wg    
│          │  subnet → table    
│          │  with multipath    ┌───────────┐
│ vprox0t0 │◄──────────────────►│  vprox0   │◄─── UDP :50227 ───► vprox0t0 (wg)
│ vprox0t1 │◄──────────────────►│  vprox0t1 │◄─── UDP :50228 ───► vprox0t1 (wg)
│ vprox0t2 │◄──────────────────►│  vprox0t2 │◄─── UDP :50229 ───► vprox0t2 (wg) ──► Internet
│ vprox0t3 │◄──────────────────►│  vprox0t3 │◄─── UDP :50230 ───► vprox0t3 (wg)     (SNAT)
└──────────┘                    └───────────┘

Backwards compatibility is preserved when the parameter is omitted or when passing --tunnels 1.