MCP server: expose the PDF and audio tools that already run in Node

Part of #57.

Goal

A small local MCP server, runnable with npx, that lets AI agents use Fizzdoc's PDF and audio engines on the user's own files.

Tool	Engine
`merge_pdf`, `split_pdf`, `extract_pdf_pages`, `delete_pdf_pages`, `reorder_pdf_pages`, `rotate_pdf`	`src/engine/pdf.ts` (qpdf)
`protect_pdf`, `unlock_pdf`, `clean_pdf` (remove metadata)	`src/engine/pdf.ts` (qpdf)
`trim_audio`, `split_audio`, `merge_audio` (MP3/M4A, no re-encoding)	`src/engine/audio.ts`

Location: a new mcp/ package in this repo, using the official MCP TypeScript SDK over stdio.
Engines: call the engines directly. The Web Worker in src/engine/worker.ts exists only to keep the browser tab responsive, and Node doesn't need it.
Inputs and outputs: inputs are absolute file paths. Outputs are written next to the input (or to an output_dir argument) and never overwrite an existing file. Each tool returns the paths it wrote.
Errors: reuse the existing error codes (PdfError, AudioError) and turn them into readable messages. Page ranges use the same syntax as the website, e.g. 1-3, 8.
Privacy: no network access at all. Please don't add dependencies that phone home.

Running npx fizzdoc-mcp (or similar) starts the server, and every tool above works on the fixtures in tests/fixtures/.
Each tool has a Vitest test that checks the output: for example, page count after a split, or duration after a trim.
A short mcp/README.md explains how to add it to an MCP client.