Consolidate `mini_swe_agent` (v1) and `mini_swe_agent_2` (v2)

[anwithk]

Context

NeMo Gym currently ships two mini-swe-agent integrations:

• mini_swe_agent - mini-swe-agent v1 (nv-mini-swe-agent fork), swegym_runner, Docker/Singularity, SWE-Gym train + SWE-bench Verified val, RL token/logprob passthrough
• mini_swe_agent_2 - mini-swe-agent v2.1.0, Gym sandbox API (OpenSandbox), SWE-bench Verified eval, richer aggregate metrics

They overlap in purpose but serve different deployment/training paths today.

Goal

Decide whether to deprecate mini_swe_agent in favor of mini_swe_agent_2, or keep both with clearer docs on when to use each.

Before deprecating v1, confirm v2 covers:

• SWE-Gym dataset + config parity
• RL training needs (token IDs, logprobs in rollout output)
• Singularity/HPC path, or documented migration to sandbox-only
• Docs/README updated (mini_swe_agent_2 not yet in root README table)

Acceptance criteria

• Written recommendation: deprecate, keep both, or merge
• If deprecating: migration guide + deprecation timeline
• If keeping both: clear "use v1 vs v2" section in docs

[cwing-nvidia]

I'd advocate that we update the README asap to direct users to v2. and then a separate effort to deprecate v1