feat(deps): CUDA 13 torch on Linux aarch64, CPU elsewhere; deterministic CUDA kernels

[ajbarea]

[tool.uv.sources] now routes torch and torchvision by platform marker: Linux aarch64 to the pytorch-cu130 index, every other platform (CI included) to pytorch-cpu. The commented cu128 opt-in it replaces had no aarch64 build. The lock resolves one version everywhere, torch 2.14.0 / torchvision 0.29.0.

set_seed calls torch.use_deterministic_algorithms(True) when CUDA is present, covered by a CUDA-only test. ROADMAP.md now scopes the replay claim to client training and tracks run-level replay: identical smoke runs ended at 0.979 and 0.721 because Flower's client sampling is unseeded over per-run random node IDs.

Verification:

• x86 CPU: make lint, make test (18 passed, CUDA test skipped), make audit, make smoke exit 0.
• TIGRIS GH200 (driver 610.43.02): CUDA 13 torch runs on the device (102 GB, bf16); deterministic mode gives bitwise-equal embedding + linear gradients across runs, with or without CUBLAS_WORKSPACE_CONFIG.
• TIGRIS GH200, torch 2.14.0+cu130, at cfaf0d197 (this branch with main merged): make lint clean; pytest 43 passed, 0 skipped, including the CUDA-only deterministic test; GPU make smoke (client-resources-num-gpus=0.2) 2 rounds, 0 failures. GPU memory rose from 0 to 3,467 MiB during the rounds and utilization peaked at 49%; nvidia-smi lists no processes on these nodes, so memory is the evidence.
• TIGRIS GH200 at ec1f87136: the CUDA test compares a float32 index_add_ over duplicate indices. With deterministic mode off, 5 runs gave 5 distinct results (max abs diff 3.7e-3); with it on, 5 identical. Full suite 43 passed, 0 skipped.

[ajbarea]

Rebased onto main to clear a uv.lock conflict. The branch was force-pushed, so a local copy needs git fetch && git reset --hard origin/feat/gpu-on-linux-aarch64 before any further work on it.

The conflict came from chore: refresh the dev toolchain (#97), which rewrote the lock for ruff/ty/pytest while this branch rewrites it for the torch index routing. pyproject.toml merged cleanly — the [tool.uv.sources] block and the dev-dependency floors do not overlap — so only the lock needed resolving, and a lock is regenerated rather than hand-merged.

Taking main's lock and re-resolving held torch at 2.13.0, because uv lock keeps an existing pin that still satisfies the constraint and torch>=2.12 does. An explicit --upgrade-package torch --upgrade-package torchvision restores what this branch intends. Both are also current: 2.14.0 is the latest on PyPI and is published on the cu130 aarch64 index and the cpu index, so the routing resolves on both legs.

Lock now carries torch 2.14.0 / torchvision 0.29.0 from here, and ruff 0.16.8 / ty 0.0.82 / pytest 9.1.1 / anyio 4.14.2 from main, with 11 cu130 and 20 cpu wheel URLs recorded.

Checked locally on x86 CPU: ruff format --check, ruff check, ty check, uv lock --check all clean, and pytest --cov=phalanx passes at 18 passed, 1 skipped — the skip being the CUDA-only deterministic test. CI is green on all required checks.

Not merging: the description names a GPU make smoke on a GH200, queued as a Slurm batch job, as the gate.