fix: decouple oa-vm from the ML training stack via lazy package imports

[abrichr]

Problem

oa-vm (VM lifecycle CLI) crashed on startup:

A module that was compiled using NumPy 1.x cannot be run in NumPy 2.0.2 ...
  from openadapt_evals.benchmarks.vm_cli import main
  ...eventually...
  from transformers import TrainerCallback

Root cause is an import-chain problem, not NumPy: oa-vm's entry point
imports openadapt_evals.benchmarks.vm_cli, which triggers the package
__init__ modules. Those eagerly did from openadapt_evals.agents import ...,
cascading into transformers/peft — the entire ML training stack — just to
start and stop VMs. Under a NumPy 2 / stale-transformers environment that
cascade hard-crashes, taking oa-vm down with it. Downgrading NumPy would be a
workaround; a VM CLI should not import the training stack at all.

Fix

Resolve the re-exported public API of openadapt_evals and
openadapt_evals.benchmarks lazily (PEP 562 module __getattr__), caching
each name as a module global on first access. The repo already used this pattern
for PolicyAgent/Azure; this extends it to the heavy re-exports.

• Importing either package is now light and pulls in no ML/vision
  dependency. agents/transformers load only when an agent-dependent name is
  actually accessed.
• Public API and object identity are unchanged: from openadapt_evals import BenchmarkAgent and oe.WAAAdapter is <direct import> both hold.

Verification

• oa-vm --help and oa-vm pool-status run cleanly (both in the broken
  anaconda env that reproduced the crash and in the uv env).
• Full public API resolves with identity preserved.
• New subprocess regression guard TestVmCliStaysLight asserts the oa-vm
  import path pulls in none of transformers/peft/torch/agents, and that
  lazy re-exports still resolve to the real objects.
• Test suite: 302 passed, 1 pre-existing unrelated failure
  (test_demo_persistence.py::...test_demo_format_and_persistence imports the
  optional openadapt_ml directly without an importorskip guard — fails on
  main too; flagged, not in scope here).

🤖 Generated with Claude Code