Add initial Qwen3.5-4B support and SGLang text-runtime patch

[ShenAC-SAC]

Summary

This PR adds initial Qwen3.5 support for the personal OpenClaw training paths:

• OpenClaw Combine
• OpenClaw RL
• OpenClaw OPD

The goal of this PR is to wire Qwen3.5 into the current OpenClaw/SLIME/Megatron stack and verify that the model can reach a real startup smoke path in the existing training/runtime flow.

Main Changes

Added Qwen3.5 integration across the current stack:

• slime/slime_plugins/models/qwen3_5.py
• slime/slime_plugins/mbridge/qwen3_5.py
• slime/slime/backends/megatron_utils/megatron_to_hf/qwen3_5.py

Added or updated Qwen3.5 launch scripts:

• slime/scripts/models/qwen3.5-4B.sh
• openclaw-combine/run_qwen35_4b_openclaw_combine.sh
• openclaw-rl/run_qwen35_4b_openclaw_rl.sh
• openclaw-opd/run_qwen35_4b_openclaw_opd.sh

Added repo-side SGLang compatibility work for Qwen3.5 rollout serving:

• slime/slime/backends/sglang_utils/qwen3_5.py
• slime/slime/backends/sglang_utils/sglang_engine.py

Added a low-resource debug-rollout startup fix for PRM-based methods:

• slime/slime/ray/placement_group.py

Dependency / Runtime Notes

Qwen3.5 support does not work with the current repo pin transformers==4.57.1.

During isolated validation, Qwen3.5 required a newer Transformers build that provides transformers.models.qwen3_5.
The validation environment used:

• transformers==5.2.0

For runtime validation, I also used a newer SGLang-related stack in Docker:

• sglang==0.5.9
• flashinfer-python==0.6.6
• sgl-kernel==0.3.21

For the reduced-GPU combine validation path, the environment also needed:

• numpy<2 (Megatron currently rejects NumPy 2.x)
• pylatexenc
• wandb

I have not updated the repo-wide dependency pin in this PR yet, because I wanted to first confirm the Qwen3.5 path can actually run through the current training/runtime stack before proposing a broader dependency change.

Validation Completed

Completed checks:

• Python syntax check for Qwen3.5 integration files
• Shell syntax check for Qwen3.5 launch scripts
• Qwen3.5 model spec consistency check
• isolated environment verification that transformers supports qwen3_5
• local checkpoint load validation for Qwen3.5-4B
• Dockerized single-GPU SGLang smoke test for the Qwen3.5 text rollout path
• reduced-GPU openclaw-combine startup smoke with PRM enabled

Single-GPU serving smoke result:

• server startup succeeded
• /health_generate returned 200
• /model_info returned 200
• /generate returned 200

Reduced-GPU combine smoke result:

• validated with a reduced-GPU --debug-rollout-only startup path
• fixed placement-group allocation so debug_rollout_only can still reserve PRM GPUs
• fixed SGLang server-arg handling so a Qwen3.5 shadow text-only checkpoint is not incorrectly forced back through language_only -> encoder_urls validation
• verified startup now progresses through:
  • Ray placement-group creation
  • RolloutManager creation
  • rollout router startup
  • PRM router startup
  • Qwen3.5 shadow text-only checkpoint preparation
  • rollout SGLang engine launch and weight loading
  • PRM SGLang engine launch and weight loading
  • /health_generate, /server_info, and /model_info
  • OpenClaw OPD proxy startup
  • policy server ready
  • final model is ready banner

This smoke run was intentionally bounded with a timeout after the stack became ready; it did not stop because of a Qwen3.5 startup crash.

Remaining Notes

Still not completed in this PR validation:

• full multi-step training verification beyond startup smoke
• broader multi-GPU validation
• repo-wide dependency pin update decision

openclaw-test was not run here. It depends on a separately running OpenClaw gateway plus an external user-model endpoint/token, which was outside the repo-local Qwen3.5 training validation setup used for this PR.

Why This PR Is Ready For Review

At this point the main repo-side goal is met:

• Qwen3.5 is wired into the existing stack
• the runtime compatibility blockers in this branch have been addressed
• the model now reaches a real reduced-resource OpenClaw Combine startup smoke path instead of failing immediately during import, conversion, or SGLang initialization

Follow-up work such as broader dependency pinning or more exhaustive training validation can still be done in subsequent PRs if preferred.