Migrate to JDI execution control

[m-dzianishchyts]

What does this PR do?

Moves notebook execution into a separate persistent executor JVM and separates what that JVM computes from what the managing JVM shows in Jupyter. The branch carries both steps:

• Executor JVM: cells run in a persistent executor JVM controlled through JDI.
• Output separation: the display path between the two JVMs is replaced by a transport-independent protocol.

Java 21 is now the minimum version.

Details

Execution

• A cell's result is now a plain-text evaluation result. It no longer carries Jupyter display data.
• The old in-process class loading, execution control, telemetry collectors and the standalone launcher are removed.
• The %time magic is simplified accordingly.
• Standard input for JShell now comes from the Jupyter input stream.

Display and update between the JVMs

• Explicit display and updateDisplay calls are encoded as validated, transferable requests. Cyclic structures and non-finite numbers inside bundles or containers are rejected before any connection is opened.
• Delivery is one request and one response per connection over a Unix-domain socket, with a bounded wait and no retries.
• Failures are reported as unavailable, rejected or uncertain. The uncertain case never claims that nothing was shown, and a healthy executor stays usable afterwards.

Helper routing

• The managing JVM and the executor JVM now use separate helper classes, and which one runs is fixed by what each JVM imports.
• Executor printf, local render, display and update are supported. Nested evaluation, programmatic magics and kernel access in the executor fail with operation-specific messages.

Compatibility

• Java 11 and 17 are no longer supported. Build and CI target 21 and 25.

Testing

• the protocol codec and validation;
• delivery and endpoint behavior, including the timeout bound;
• helper routing and render characterization;
• magic dispatch;
• dependency boundaries between the evaluator, protocol and Jupyter layers.
• the display round trip;
• failures leaving the next cell runnable;
• a stalled manager;
• endpoint lifecycle.