Feat/refactoring: migration to Transformers v5, removing custom MoE backends, and misc. improvements

[fedebotu]

Overview

This PR replaces the custom MoE backend system (vLLM, SGLang, FlashInfer, HF -- also noting that newer version of these were quite buggy) with the recently released transformers v5's built-in Qwen3MoeExperts and Qwen3MoeTopKRouter. This drops ~250 lines of backend-specific code and lets transformers handle kernel dispatch automatically.

Additionally, several changes to the overall QOL, including linting, bug fixes, and documentation, have been added.

The older version that included the backends code will be left open at the backends brach for future reference.

Major Changes

Model (rnd/modeling_rnd.py)

• Rewrote RND1SparseMoeBlock to use Qwen3MoeExperts + Qwen3MoeTopKRouter instead of manually routing tokens through per-expert MLPs
• Removed all vLLM/SGLang/FlashInfer imports, weight-packing logic, and backend selection
• Registered rnd1 in transformers' _MODEL_TO_CONVERSION_PATTERN so per-expert checkpoint weights are automatically fused into the 3D tensor format during loading
• Added rotary embedding inv_freq recomputation in from_pretrained (these buffers are non-persistent and not stored in safetensors)
• Simplified RND1DecoderLayer and RND1Attention by removing backend-conditional RMSNorm class selection
• Replaced _init_weights with a no-op (weights always comAdditionally, several changes to the overall QOL including linting, bug fixes, and documentation have been added.e from a checkpoint)

Config (rnd/configuration_rnd.py)

• Removed moe_backend parameter
• Switched RoPE config to the v5 rope_parameters format

Demo script (demo_rnd_generation.py)

• Replaced --moe_backend with --experts-implementation (optional, auto-detected when not set)
• Normalized all CLI args to kebab-case (--top-k, --num-steps, etc.) for consistency with other projects

Project setup

• Bumped minimum: transformers>=5.0.0, torch>=2.8
• Removed optional deps for vllm, sglang, flashinfer
• Reduced ruff line-length from 120 to 100, added bugbear/simplify/isort rules
• Cleaned up pre-commit config (removed generic hooks, updated ruff, fixed uvx commands)
• Removed unused SVG asset