Fix master routing tasks to instances with dead runners

[mlpy0]

Problem

I hit this running tensor-parallel inference (Qwen3.5-35B, JACCL across 2 Mac Minis) under heavy load — 24 concurrent long-prompt requests. One runner crashed with SIGABRT from Metal GPU synchronization (mlx::core::Event::wait() → IOSurfaceSharedEvent waitUntilSignaledValue:timeoutMS:). The EXO process stayed alive (HTTP API on port 52415 responding, /state returning data) but the runners were dead:

Runner 84fa9aa5: RunnerFailed — "Terminated (signal=6 (Abort trap: 6))"
Runner 0131b461: RunnerShuttingDown  (killed by plan._kill_runner after sibling failed)

After this, every /v1/chat/completions request hung indefinitely. No error, no timeout — just an open connection that never gets a response.

Root Cause

Three issues in the master compound to cause the hang:

1. Command processor routes to dead instances

The TextGeneration handler in _command_processor selects the instance with the fewest active tasks. After a crash, the dead instance has 0 tasks (all previous tasks got ErrorChunk from the supervisor), so it gets picked first. The worker receives the task but plan() short-circuits to _kill_runner — the TextGeneration task is never executed. The API's _token_chunk_stream blocks forever waiting for chunks.

2. "No instance found" error is silently swallowed

The except ValueError at the end of _command_processor catches "No instance found for model X" and logs a warning, but doesn't send any event back to the API. If the model genuinely isn't loaded (or all instances have dead runners), the API creates a channel and waits for chunks that never arrive.

3. Zombie instances persist after runner crash

_plan() only deletes instances when nodes disconnect from the topology. After a runner crash, the node is still connected (EXO's main process is alive, HTTP API responds, libp2p peer is up). The instance stays in state.instances forever with dead runners — a zombie that matches model lookups but can never serve requests.

Fix

All changes in src/exo/master/main.py:

1. Skip instances with failed runners — Before counting an instance as available for TextGeneration routing, I check if any of its runners are in RunnerFailed state. I use the negative check (RunnerFailed) rather than a positive check (RunnerReady/RunnerRunning) because during normal startup, runners aren't in state.runners yet — a positive check would break task queuing while models are loading.

2. Send ErrorChunk on "no instance found" — In the except ValueError block, I send a ChunkGenerated(ErrorChunk) event for TextGeneration commands so the API gets a proper error response instead of hanging. This follows the same pattern used in runner_supervisor.py for runner crash errors.

3. Delete instances with failed runners in _plan() — Alongside the existing node-disconnect cleanup, I also delete instances where any runner is RunnerFailed. I use the for/else pattern to avoid double-deleting when both checks would trigger.

Testing

All existing tests pass (331 passed, 1 skipped).