Add dense Megatron LoRA-XS support

[kalectory]

• Adds dense Megatron LoRA-XS with frozen SVD factors and a trainable rank-by-rank core, building on the PiSSA distributed initialization path.
• Keeps algorithm behavior in lora_xs.py; the shared worker only selects the PEFT class and invokes its optional export context.
• Adds offline step-zero initialization, Qwen3-0.6B GSM8K example scripts, and focused coverage for factorization, TP sharding, gradient isolation, fused-forward handling, and export restoration. Stacked on Add PiSSA initialization for Megatron LoRA adapters NovaSky-AI/SkyRL#2069.

Testing

NVTE_FLASH_ATTN=0 uv run --isolated --extra dev --extra megatron pytest tests/backends/skyrl_train/gpu/gpu_ci/megatron/test_lora_xs.py -q

9 passed; verifies the focused LoRA-XS initialization and export behavior.

uv run --isolated --with pre-commit pre-commit run --files skyrl/backends/skyrl_train/workers/megatron/lora_xs.py skyrl/backends/skyrl_train/workers/megatron/megatron_worker.py tests/backends/skyrl_train/gpu/gpu_ci/megatron/test_lora_xs.py

Formatting, lint, and secret checks passed.

On 8x H200 with Qwen3-0.6B/GSM8K, offline initialization plus one 16-trajectory GRPO batch completed policy training, saved global_step_1, and synchronized the updated weights to vLLM.