Add Agent Plugins specification support

## Summary

Add support for the Agent Plugins open specification so `tiny-coding-agent` can discover and consume portable agent plugins across compatible clients.

Reference: https://agent-plugins.org/

## Motivation

Agent Plugins provides a common packaging format for reusable agent capabilities, including skills and MCP server configuration. Supporting the specification would let `tiny-coding-agent` reuse plugins built for the broader agent ecosystem instead of requiring project-specific integration.

This also fits the project's goal of keeping the coding agent small while allowing capabilities to be extended externally.

## Proposed scope

### Phase 1 — Plugin discovery and manifest

- [ ] Discover plugin directories
- [ ] Parse and validate `plugin.json`
- [ ] Support the Agent Plugins schema/version field
- [ ] Expose discovered plugin metadata to the agent runtime
- [ ] Provide useful validation/error messages for invalid plugins

### Phase 2 — Skills

- [ ] Discover `skills/` declared/provided by a plugin
- [ ] Load `SKILL.md` instructions using the existing skills mechanism where possible
- [ ] Keep plugin skills isolated/namespaced to avoid collisions
- [ ] Define precedence when local and plugin-provided skills share a name

### Phase 3 — MCP

- [ ] Read plugin `mcp.json`
- [ ] Register configured MCP servers with the existing tool/MCP layer
- [ ] Handle environment/configuration requirements without embedding secrets in plugins
- [ ] Surface MCP startup/configuration errors clearly

### Phase 4 — CLI / UX

Consider a minimal interface such as:

```bash
tiny-agent plugin list
tiny-agent plugin inspect
tiny-agent plugin validate
```

Installation/registry support can remain out of scope initially; loading plugins from configured local directories is enough for the first implementation.

## Design principles

- Follow the Agent Plugins specification rather than inventing a custom plugin format.
- Keep the core agent runtime independent from individual plugins.
- Reuse the existing skills and MCP implementations instead of introducing parallel systems.
- Prefer progressive loading so plugin metadata can be discovered without loading every skill/tool into context.
- Treat plugin configuration and executable MCP servers as untrusted input.

## Acceptance criteria

A minimal example plugin containing:

```text
example-plugin/
├── plugin.json
├── skills/
│ └── example/
│ └── SKILL.md
└── mcp.json
```

can be discovered by `tiny-coding-agent`, validated, and its skill/MCP capabilities made available to the agent runtime.

## References

- Agent Plugins: https://agent-plugins.org/
- Specification: https://agent-plugins.org/specification
- Plugin author guide: https://agent-plugins.org/plugin-authors

The issue names the existing skills mechanism, tool/MCP layer, and TypeScript CLI/runtime, but no repository files or tests. Start by locating those entry points and reviewing the Agent Plugins specification; done means a local example plugin can be discovered, validated, and expose its skills and MCP capabilities through the runtime.

Written by the indexing model from the issue text.