Add support for showing variables

**Problem:**

`get_issue_details` returns stack frames with file, line, function, and source code, but not local variable values — even when Sentry has captured them. Without variable context, the assistant can't actually diagnose the issue — it has to punt back to the human to check the UI. That defeats the purpose of the MCP integration.

I ran into this today. We had an `EncodeError: Object of type str is not JSON serializable` in a Celery/Kombu stack. In the Sentry web UI, I can see the value of `o` at the crash site (`json/encoder.py:180`) — it was a `SimpleLazyObject`. That's the whole answer to "what was being encoded?" But the MCP tool only gave me frame metadata, so I had to leave the AI workflow and go look at the UI myself.

**Proposed solution:**

Rather than bloating every `get_issue_details` response with variables from every frame, add a separate tool like `get_frame_variables(issue_id, event_id, frame_index)` that returns the local variables for a specific stack frame. This keeps the common case token-efficient while still letting the assistant drill into the frame that matters.

Another option to save on # of requests might be to include variable data only for the top/bottom frame of each exception by default, since that's where the most useful context usually lives.