refactor: move logging config to CLI entrypoint and redact payloads (A5)

[ergut]

This closes A5, the last remaining item from the architecture review.

What changed

• Importing mcp_logseq.server no longer configures logging. The old import-time block forced the root logger to DEBUG and unconditionally wrote ~/.cache/mcp-logseq/mcp_logseq.log. A subprocess test now guarantees the import leaves the root logger untouched.
• Logging is configured in the CLI entrypoint instead. _setup_logging() runs at the start of main() and covers both transports: level comes from LOGSEQ_LOG_LEVEL (default INFO), logs go to stderr, and a log file is written only when LOGSEQ_LOG_FILE is set. Bad config never crashes the server: an invalid level falls back to INFO with a warning, an unopenable file degrades to stderr only.
• Tool dispatch logging is redacted. The call_tool body moved into a module-level _dispatch_tool_call() coroutine, and logs now carry only the tool name, sorted argument keys, and the result item count. Page and block content no longer lands in plaintext logs, which matters given the ACL layer's intent.
• Two extra leaks found in review are also closed: a logger.info in logseq.py that dumped full page property values now logs keys only, and at LOGSEQ_LOG_LEVEL=DEBUG the MCP SDK logger stays capped at INFO so inbound JSON-RPC payloads are not written to the log file.
• README documents the two new env vars; CHANGELOG gets a "potentially breaking" entry: the default level drops from DEBUG to INFO and the log file is now opt-in, so anyone relying on ~/.cache/mcp-logseq/mcp_logseq.log must set LOGSEQ_LOG_FILE.

Testing

Full suite: 643 tests passing (13 new in tests/unit/test_logging.py covering level parsing, file handler opt-in, import purity, and redaction).

Design spec and implementation plan are included under docs/superpowers/.