run_mcts spins forever when the data-loader node can't be created, writing GBs of log

Found while smoke-testing the 1.0.0 release from PyPI. **Not a 1.0.0 regression** — the code is identical at 0.2.2 (`3e976267`). Any durable model failure at the data-loader step triggers it; #78 just made it easy to reach.

## Symptom

A run whose model calls fail persistently does not exit. It emits, at full speed with no backoff:

```
SAMPLED 1 NODE(S) FOR EXPANSION: ['0_0']
==========================
(1/1): EXPANDING NODE 0_0
No new experiment generated for node 0_0. Skipping this iteration.
```

Measured: **3.9 GB of log in 90 seconds** (~204M lines), of which exactly 3 lines were the underlying error. The real failure is unfindable, and an unattended run will fill the disk.

## Mechanism

[`mcts_utils.py:271`](../blob/v1.0.0/packages/autodiscovery/src/autodiscovery/mcts_utils.py#L271):

```python
# If the data loader node hasn't been executed, return the root. This is not run in batch mode.
if len(nodes_by_level[1]) == 0:
return [root]
```

The root's only job is to produce the data-loader node at level 1. If the model call that generates it fails, `nodes_by_level[1]` stays empty, so `select_nodes` returns `[root]` **unconditionally and forever**. This early return never consults [`has_untried_experiments()`](../blob/v1.0.0/packages/autodiscovery/src/autodiscovery/mcts.py#L194), so the `allow_generate_experiments = False` guard set at [`run.py:541`](../blob/v1.0.0/packages/autodiscovery/src/autodiscovery/run.py#L541) has no effect on this path.

Meanwhile the outer loop at [`run.py:490`](../blob/v1.0.0/packages/autodiscovery/src/autodiscovery/run.py#L490):

```python
while n_sampled < total_to_sample:
```

never advances, because the node is never committed — the failure path `return None`s at `run.py:546` without incrementing `n_sampled`. The `if not next_nodes: break` escape at `run.py:503` is unreachable, since `select_nodes` always returns a non-empty list. No iteration cap, no consecutive-failure counter, no sleep.

## Triggers

Anything that makes data-loader generation fail durably rather than transiently:

- misconfigured or missing Vertex project/location (#78)
- expired or insufficient credentials
- quota exhaustion, or a model the project has no access to
- a model whose responses never parse into an experiment

## Fix

Two independent guards, either of which stops the flood; both are cheap:

1. **Bound no-progress iterations in `run_mcts`.** Track consecutive iterations that commit no node and abort with the last underlying error once past a small threshold. This catches every cause, not just the data-loader case.
2. **Stop returning `[root]` unconditionally.** Once the root has failed to produce a data loader and has no untried experiments, `select_nodes` should return `[]` so the existing `if not next_nodes: break` fires.

(1) alone is sufficient and is the more general fix. Doing both means the loop terminates even if a future selection method reintroduces an unconditional return.

## Acceptance

- [ ] A run whose experiment generation always fails terminates with a non-zero exit and surfaces the underlying error, rather than looping.
- [ ] Log output for that run is bounded — no unbounded repetition of the same four lines.
- [ ] A regression test drives `run_mcts` with an experiment generator that always fails and asserts termination.

Related: #34's unchecked *"Enable public repo integration tests"* item — this is precisely the class of bug that only appears when a real model call fails durably, so no current test can catch it.