cuda: Module.holds, whether a loaded image defines a device global

[plotarmordev]

What this changes

Module.holds(name) reports whether a loaded image defines the device global name, so the host can check a marker
before relying on what it promises (part of #914). Any driver failure, a missing name included, answers false
rather than an error.

Drama 3 uses it to launch with programmatic dependent launch only on images that carry its marker.

Receipt

• Environment, with the TensorFold commit, MLX or PyTorch version, machine and GPU, checkpoint and revision:
  • Commit ab7659dae7e6, on main c7137bb3. No MLX or PyTorch.
  • An NVIDIA GB10 build host on Ubuntu 24.04.4 LTS aarch64 (Linux 6.17), with Zig 0.17.0 and nvcc 13.0.88.
  • No checkpoint is involved.
• Exactness, from tools/bench_concurrent.py --alone --serial or token_sha against "draft": false: not run. The
  change adds a query that nothing calls yet.
• Decode speed before and after, from tools/bench_openai.py: n/a. Nothing on the decode path calls it.
• Prompt speed before and after: the cold first request's ttft_s from tools/bench_cache.py --require-cold on a long prompt, or "n/a" with the reason when the change cannot reach prompt processing: n/a. Nothing on the
  prompt path calls it.
• Tests run, with pass, fail and skip counts:
  • CUDA_VISIBLE_DEVICES= zig build test-cuda-host --summary all from an empty local cache: 27/27 steps, 466 passed, 0 failed, 0 skipped.
  • The new test, "holds answers whether the image defines a global, and false on any driver failure", is CPU-only.
    It drives holds through a fake driver table for three cases:
    • a defined global (true);
    • a missing name, CUDA_ERROR_NOT_FOUND (false);
    • another driver error, CUDA_ERROR_INVALID_CONTEXT (false).
  • A holds that treats only a missing name as false fails the test: 83 passed, 1 failed (that test). Before the change the test does
    not build, because holds does not exist.
  • zig build native -Dnvcc=/usr/local/cuda/bin/nvcc --summary all: 42/42 steps.
  • python3 tools/zig/lean_check.py: 527 problems, the same as main's 527 (tools/zig/lean_new.py: 0 new).
  • zig fmt --check build.zig zig: clean.
• Not run, and why:
  • A loaded image on a GPU. Its device check, tf-cuda-test drama-pdl, arrives with its first caller, Drama 3's
    programmatic dependent launch.
  • Metal: not applicable; the change is in the CUDA runtime.

Checklist

• Every token still goes through the lane rounds. No serial path, and nothing that needs drafts off. (No token
  path is touched.)
• Drafted output equals "draft": false, a resumed prompt equals a fresh one, and concurrent equals solo. (Not
  run; no output path is touched.)
• No precision traded for speed. If the bits change, the description says which sums change and why. (No
  arithmetic changes.)
• Prompt processing is no slower than the last release. (Not measured; nothing on the prompt path calls it.)
• The description says which platforms I ran, Metal M1 to M5 and CUDA, and which I could not.
• New tests fail before the change, pass after it, and skip cleanly without their dependency. (It needs no GPU.
  The fail-closed case fails on a holds that only treats a missing name as false.)
• Comments and docstrings are one line. No measurements or history in the source.
• No personal data, machine names, internal hosts or local paths. No AI attribution lines.