Rotation tasks for v0.15.0

[chensuyue]

[lkk12014402]

rotation features:

1. layer-wise rotation, which is compatible with layer-wise quantization, for larger LLM
2. support moe
3. verification on xpu
4. verify spinquant tuning recipe: spinquant block-wise training (trainable rotation matrix) + autoround block-wise tuning

[lkk12014402]

rotation features:

1. layer-wise rotation, which is compatible with layer-wise quantization, for larger LLM
2. support moe
3. verification on xpu
4. verify spinquant tuning recipe: spinquant block-wise training (trainable rotation matrix) + autoround block-wise tuning

1. https://github.com/lkk12014402/auto-round/tree/quarot-quant-block-wise-vllm

[lkk12014402]

blockwise rotation: #2087

[lkk12014402]

xpu verification: #2094

[lkk12014402]

MoE support: https://github.com/lkk12014402/auto-round/tree/rotation_moe_test

I am currently evaluating the accuracy impact of rotation on Qwen3.6-35B-A3B.