fix: distinguish AddrInUse from transient IPv6 bind failures in port checker

[sadjow]

Summary

Fixes false-positive "Port X is already in use" errors when --strict-ports is enabled and the nix evaluation cache is invalidated.

• bind_no_reuse now returns the actual io::Error instead of discarding it as ()
• reserve_port only treats AddrInUse errors from the IPv6 bind as genuine port conflicts
• Non-conflict IPv6 failures (transient OS state on macOS, permission errors) fall through to IPv4-only mode instead of failing the entire port check

Problem

On macOS, after a clean devenv up shutdown (all processes confirmed dead, lsof shows no listeners), restarting immediately can fail with "Port 6380 is already in use" but only when:

1. The nix eval cache is invalidated (e.g. a file changed)
2. --strict-ports is enabled

The port is not actually in use. lsof confirms nothing is listening and get_process_using_port finds no PID. The issue is that bind_no_reuse on [::1]:port can fail for non-AddrInUse reasons on macOS, and the previous code assumed any IPv6 bind failure meant the port was occupied.

Test plan

• Added test_bind_no_reuse_returns_addr_in_use_error to verify error kind propagation
• Added test_reserve_port_succeeds_when_ipv4_free to verify strict mode works on free ports
• Added test_strict_mode_fails_only_on_genuine_conflict to verify real conflicts are still detected
• All existing port allocator tests should continue to pass