[evals] Rewrite evals.py, drop legacy scaffolding, levanter->function

Wires the new typed API into the user-facing layer and cleans up the
pre-OpenAI-HTTP scaffolding that no longer has callers.

experiments/evals/evals.py
- Every helper now builds typed (ModelDeployment, LmEvalRun|HarborRun)
  pairs; step runners use @Remote(resources=...) v2-Fray wrappers instead
  of the v1 launch_evaluate_with_ray path.
- engine_kwargs param split into explicit deployment_kwargs (vLLM server
  flags) + extra_model_args (lm-eval client knobs) + batch_size.
- evaluate_harbor supports external-API mode (model_path is None) by
  building a RunningModel with LITELLM_PROVIDER_URL.
- default_eval stays on Levanter shim until #4828 lands (step 10).

experiments/evals/engine_configs.py
- DEFAULT_LM_EVAL_MODEL_KWARGS split into DEFAULT_VLLM_DEPLOYMENT_KWARGS
  + DEFAULT_LM_EVAL_EXTRA_MODEL_ARGS, matching the new API shape.

Callers updated
- run_base_model_evals.py, exp_evalchemy_eval.py,
  exp_evalchemy_eval_reproduce_openthoughts.py: engine_kwargs dicts
  split into deployment_kwargs + extra_model_args + batch_size args.

Levanter evaluator
- LevanterLmEvalEvaluator class collapsed into a run_levanter_lm_eval()
  function. No Evaluator ABC / ModelConfig coupling. Scheduled for full
  deletion in step 10 (gated on #4828).

Deleted
- marin.evaluation.run (draccus CLI, evaluate(config), EVALUATORS dict,
  _impute_model_config, _to_v1_resource_config adapter, _normalize_model_path)
- marin.evaluation.evaluators.evaluator (Evaluator ABC, ModelConfig,
  Dependency, v1-Fray launch_evaluate_with_ray free function)
- marin.evaluation.evaluators.simple_evaluator (the "debug" mapping
  was only reachable via the deleted run.py:main)
- marin.evaluation.evaluators.levanter_tpu_evaluator (base class for
  the removed LevanterLmEvalEvaluator)
- marin.evaluation.evaluation_config.EvaluationConfig
- marin.inference.vllm_server.resolve_model_name_or_path /
  _maybe_enable_streaming (legacy ModelConfig shims)

Tests
- tests/evals/test_lm_eval.py: @tpu_ci guard updated to the new helper
  signature; test_lm_eval_harness_levanter dropped (duplicate Tier C —
  default_eval still exercises the Levanter path via other experiments).
- tests/evals/test_evals_helpers.py: new migration tests — each helper
  builds the expected (ModelDeployment, run-config) pair, Harbor
  external-API vs local-vLLM modes, parameterized step-suffix extraction.

Known out-of-scope issue, flagged for separate handoff: the evalchemy
commit pin 010412c set on main in PR #3690 is not reachable on either
teetone/evalchemy or mlfoundations/evalchemy, so `git checkout 010412c`
at runtime fails. Not introduced by this PR.

Author: yonromai

experiments/evals/engine_configs.py

@@ -1,8 +1,16 @@
 # Copyright The Marin Authors
 # SPDX-License-Identifier: Apache-2.0
 
-"""Engine configuration for vLLM used for evals."""
+"""Engine + run defaults for vLLM-backed evals.
 
-DEFAULT_VLLM_ENGINE_KWARGS = {"max_model_len": 4096}
+Splits today's single `DEFAULT_LM_EVAL_MODEL_KWARGS` bag into two halves that
+target the post-#4827 types:
 
-DEFAULT_LM_EVAL_MODEL_KWARGS = {**DEFAULT_VLLM_ENGINE_KWARGS, "max_gen_toks": 4096}
+- `DEFAULT_VLLM_DEPLOYMENT_KWARGS`: vLLM server flags. Feeds `ModelDeployment.engine_kwargs`.
+- `DEFAULT_LM_EVAL_EXTRA_MODEL_ARGS`: per-request / lm-eval client knobs.
+  Feeds `LmEvalRun.extra_model_args` as pre-formatted `k=v` strings.
+"""
+
+DEFAULT_VLLM_DEPLOYMENT_KWARGS: dict = {"max_model_len": 4096}
+
+DEFAULT_LM_EVAL_EXTRA_MODEL_ARGS: tuple[str, ...] = ("max_gen_toks=4096",)