[ci] build the PPU image from this repo

[tiankongdeguiji]

Build the PPU image from this repo instead of an out-of-tree addons checkout, and pick up the PPU artifacts' new home.

The PPU image was built outside the repo, so it drifted from requirements/ whenever the release pins moved and could not be reproduced from a checkout. docker/Dockerfile.ppu now installs the PPU torch/triton/faiss/fbgemm builds from the vendor pypi index on top of the vendor PPU base image, then installs requirements.txt like the other images do. The torch==2.11.0 / fbgemm-gpu==1.6.0 / torchrec==1.6.0 pins in requirements/runtime.txt are satisfied by the PPU builds (their local version +v0.1.0.ppu2.1.1 matches the pin), so no non-PPU wheel is pulled in.

Changes:

• docker/Dockerfile.ppu: base image ppu:v2.1.1-cuda13.0-ubuntu24-py312, PPU wheels from the vendor index, mirrors.aliyun.com for everything else. Drops the PIP_INDEX_URL + sitecustomize.py / aiextpypi.pth STS shim the old out-of-tree Dockerfile carried; it is no longer needed.
• scripts/build_docker_ppu.sh / scripts/promote_docker_ppu.sh: build and promotion for the single PPU tag. Kept separate from build_docker.sh / promote_docker.sh because the PPU image is built and validated on its own cadence and may lag the cpu/cu* images; it also never takes the <tag> / latest aliases.
• .github/workflows/unittest_ppu_ci.yml: run the PPU lane on the freshly built image (switched back to tzrec-devel:1.2-ppu after promotion).
• docs/source/develop.md: document the PPU image build.
• Version bumped to 1.2.24.

Test Plan

• Built the image from this branch and smoke-tested it off-device: torch 2.11.0+v0.1.0.ppu2.1.1, triton 3.6.0+v0.2.0.ppu2.1.1, faiss 1.14.1+v0.1.0.ppu2.1.1, fbgemm_gpu 1.6.0+v0.1.0.ppu2.1.1, torchrec 1.6.0; torch, triton, faiss, fbgemm_gpu, torchrec, torchmetrics, graphlearn, pyfg, tensorboard, sklearn all import.
• Confirmed in the build log that pip install -r requirements.txt reports torch/fbgemm-gpu/torchrec as already satisfied, i.e. the shared requirements files do not re-resolve a non-PPU torch.
• Validated cache-cold (--no-cache): an earlier build had cached a layer from when the vendor index still proxied the upstream fbgemm_gpu wheel, which loads against real CUDA but not the PPU CUDA shim (undefined symbol: __cudaGetKernel, version libcudart.so.13). The vendor index now serves only PPU artifacts for that name, and the cold build resolves fbgemm_gpu-1.6.0+v0.1.0.ppu2.1.1.
• PPU unit-test CI lane (scripts/ci/ci_test_ppu.sh, full tzrec/tests/run.py) on the built image.