Qwen3.5 Automodel path is slower than MCore because flash-linear-attention (FLA) is not installed by default

# Summary

On the Automodel (FSDP2) path, Qwen3.5 models run measurably slower than the MCore path because `flash-linear-attention` (FLA) is not installed in the Automodel worker venv by default. Several kernels in `nemo-automodel` dispatch to FLA when it is importable and fall back to slower implementations otherwise.

The root-cause fix belongs in `nemo-automodel` and is tracked separately — see https://github.com/NVIDIA-NeMo/Automodel/pull/1894, which moves FLA out of the dev-only dependency group so downstream installers (including NeMo-RL's `automodel` extra) pick it up transitively via `nemo-automodel[moe]`.

This issue tracks the NeMo-RL side:

1. Document the symptom so users picking the automodel path for Qwen3.5 (dense or MoE) are aware.
2. Bump `3rdparty/Automodel-workspace/Automodel` ; after that, no manual workaround should be needed.

# Symptoms

Same workload (Qwen3.5-35B-A3B-Base, SFT, BF16, 4K seq, 2n8g), three setups:

- Yellow: MCore path
- Red: Automodel + FLA installed
- Blue: Automodel without FLA (default today)

# Affected recipes (NeMo-RL side)

Automodel-path recipes for Qwen3.5 dense and MoE, e.g.:

- `examples/configs/recipes/llm/grpo-qwen3.5-35ba3b-2n8g-automodel-ep16.yaml`
- `examples/configs/recipes/llm/grpo-qwen3.5-35ba3b-dapo-4n8g-automodel.yaml`
- `examples/configs/recipes/vlm/vlm_grpo-qwen3.5-35ba3b-geo3k-2n8g-automodel-ep16.yaml`

(`-megatron` recipes use MCore kernels directly and are not affected.)

Also note: CP on Qwen3.5 MoE via the automodel path explicitly requires FLA today — see the `import fla` guard in `nemo_rl/models/automodel/setup.py`.

# Resolution criteria

- [x] Upstream: https://github.com/NVIDIA-NeMo/Automodel/pull/1894 merges (moves FLA to optional extras so `nemo-automodel[moe]` pulls it in by default).
- [ ] NeMo-RL: bump `3rdparty/Automodel-workspace/Automodel` to a commit that includes #1894; after that, the Automodel worker venv will install FLA automatically and this issue can be closed.