(feat): add ROLL PPU dockerfile with SAIL pypi source, full megatron-core and FLA for Qwen3.5

[SendoRay]

Summary

• Add docker/Dockerfile.torch2100.vllm.ppu for PPU hardware: SAIL internal PyPI source, pinned PPU-custom torch/vllm/flash-attn builds via PIP_CONSTRAINT
• Install full megatron-core (via mcore_adapter) to replace the base image's stripped stub
• FLA for Qwen3.5 hybrid attention: back up PPU kda shims, install fla-org build, restore shims
• Add docker/constraints_ppu.txt and requirements_torch2100_vllm_ppu.txt (transformers>=5.2.0 for qwen3_5)

Test plan

• Build image and verify import checks in Dockerfile pass (torch/vllm/roll, transformers qwen3_5, megatron DDP + mcore_adapter, fla modules + kda shims)

[CLAassistant]

All committers have signed the CLA.