Add custom paths for qwen3 and 3.5 dense

[faresobeid]

Doing this to support quack rms norm and custom selectice AC targets. Also in general cleaner to have custom paths for models we care about and serve even if non-MoE. Especially as we later adopt fp8/fp4 kernels to use

---

Note

Medium Risk
Adds new custom model implementations and changes model-config selection for Qwen3.5, which can affect training/inference correctness and attention masking/position-id behavior for these models. Main risk is regressions in Qwen3/Qwen3.5 loading and attention backends (SDPA/Flash/ring attention) rather than broader system impact.

Overview
Adds custom PrimeRL dense implementations for Qwen3 and text-only Qwen3.5, and wires them into AutoModelForCausalLMPrimeRL so impl=custom/auto-selection can instantiate these models.

Updates model loading to force Qwen3.5 text-only config when not doing VLM training (switching from composite config to text_config while preserving _attn_implementation and _name_or_path). Extends substitute_ring_attn to patch ring-attention _compute_attention for the new Qwen3/Qwen3.5 FlashAttention classes.

^{Written by Cursor Bugbot for commit 7d59e25. This will update automatically on new commits. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}