Describe the bug
When running apm install <new-package> with --trust-transitive-mcp, all transitive MCP servers are reinstalled/reconfigured even if they are already configured from a previous install. The skip logic (check_servers_needing_installation) does not apply to self-defined (transitive) MCP servers — only to registry-based servers.

To Reproduce

1. Have a project with existing APM + MCP dependencies already installed
2. Run apm install <new-package> --trust-transitive-mcp to add a new package
3. Observe that all transitive MCP servers are reconfigured even though they were already installed

Expected behavior

• MCP servers that are already configured should be skipped with ✓ <name> (already configured) — the same behavior that exists for registry-based servers.
• Only new or changed MCP servers should be installed/reconfigured.

Root Cause (analysis)
The check_servers_needing_installation() method in registry/operations.py is only invoked for registry-based MCP deps. Self-defined (transitive) MCP servers follow a separate code path in _install_mcp_dependencies() that does not check whether the server is already configured before reinstalling it.

The registry path (~L2835 in cli.py) correctly calls:

servers_to_install = operations.check_servers_needing_installation(target_runtimes, valid_servers)

But the self-defined/transitive server path does not perform an equivalent check.

Environment (please complete the following information):

• OS: macOS
• Python Version: 3.12
• APM Version: latest (main branch)

Logs

apm install xxxxx/xxxxx/general/demo-playbook --trust-transitive-mcp

Validating 1 package(s)...
✓ xxxxx/xxxxx/ - accessible
Added xxxxx/xxxxx/ apm.yml
Updated apm.yml with 1 new package(s)
Installing dependencies from apm.yml...
Installing APM dependencies (11)...
Using apm.lock (10 locked dependencies)
✓ github.com/xxxxxx/xxxxx/ (cached)
  └─ 3 prompts integrated → .github/prompts/
  └─ 1 skill(s) integrated → .github/skills/

Generated apm.lock with 1 dependencies
Installed 0 APM dependencies
┌─ MCP Servers (3)
│  ⬇️  atlassian (self-defined, http)
│     └─ Configuring for Copilot, Vscode...
  Installing atlassian...
  ✓ atlassian
│  ⬇️  zephyr (self-defined, http)
│     └─ Configuring for Copilot, Vscode...
  Installing zephyr...
  ✓ zephyr
│  ⬇️  github (self-defined, http)
│     └─ Configuring for Copilot, Vscode...
  Installing github...
  ✓ github
└─ Configured 3 servers

All 3 MCP servers were already configured from a prior apm install but were reinstalled anyway.

Additional context
This also affects the failure case: when apm install <bad-pkg> fails validation, the command still falls through and reinstalls all existing MCP servers (exits 0). Ideally, when the only requested package fails, the command should bail out or no-op instead of triggering a full MCP reinstall.