marin: port export/levanter_checkpoint.py to fray v2 (#4980)

🤖

## Summary

Ports `lib/marin/src/marin/export/levanter_checkpoint.py` from
`fray.v1.cluster.*` to `fray.v2`, mirroring the pattern in
`lib/marin/src/marin/evaluation/log_probs.py` (already on v2). Public
API — `ConvertCheckpointStepConfig`, `convert_checkpoint_to_hf`,
`convert_checkpoint_to_hf_step` — is unchanged.

This is Bucket 2 PR-B in the Ray-removal roadmap. Part of
[#4453](#4453).

## Why medium-risk

This file backs the Marin HF model-release pipeline
(`experiments/tootsie/exp_1246_upload_datasets.py` and
`experiments/tootsie/exp1984_convert_32b_phases.py`). The port preserves
every public-facing signature; only the job-submission internals change.

## Test plan

- [x] `./infra/pre-commit.py --all-files --fix` — green.
- [x] `uv run pyrefly` — no new errors vs `origin/main` baseline.
- [x] Import check: `marin.export` + both tootsie experiments import
cleanly.
- [x] Any `marin.export`-importing tests pass.
- [x] No `fray.v1` references remain under
`lib/marin/src/marin/export/`.

Author: yonromai

lib/marin/src/marin/export/levanter_checkpoint.py

@@ -7,14 +7,15 @@
 from typing import Any
 
 import levanter.infra.cli_helpers
-from fray.v1.cluster import (
+from fray.v2 import current_client
+from fray.v2.types import (
     CpuConfig,
     Entrypoint,
-    EnvironmentConfig,
+    GpuConfig,
     JobRequest,
     ResourceConfig,
     TpuConfig,
-    current_cluster,
+    create_environment,
 )
 from levanter.checkpoint import discover_latest_checkpoint
 from levanter.compat.hf_checkpoints import RepoRef
@@ -109,16 +110,21 @@ def _run_with_lockfile():
     if isinstance(config.resources.device, TpuConfig):
         assert config.resources.replicas == 1, "Export currently works on single slices at present."
 
+    extras: list[str] = []
+    if isinstance(config.resources.device, TpuConfig):
+        extras.append("tpu")
+    elif isinstance(config.resources.device, GpuConfig):
+        extras.append("gpu")
+
+    client = current_client()
     job_request = JobRequest(
         name="convert-checkpoint-to-hf",
         entrypoint=Entrypoint.from_callable(_run_with_lockfile),
         resources=config.resources,
-        environment=EnvironmentConfig.create(env_vars=env),
+        environment=create_environment(env_vars=env, extras=extras),
     )
-
-    cluster = current_cluster()
-    job_id = cluster.launch(job_request)
-    cluster.wait(job_id, raise_on_failure=True)
+    job = client.submit(job_request)
+    job.wait(raise_on_failure=True)
 
 
 def convert_checkpoint_to_hf_step(