fix(cuda-macros,codegen): keep all-generic library .oxart via #72 anchor

[dundysm]

Summary

Fixes #1365.

An all-generic #[cuda_module] in a library that monomorphizes locally still embeds .oxart with an artifact anchor, but the macro skipped the #72 rlib keep-alive under the assumption that generics never produce an artifact in the defining crate. Nothing referenced the anchor, so the linker dropped the archive member and load_all_ptx_bundles_merged returned NoModules.

Adding one unused concrete kernel "fixed" it only because concrete kernels restored the keep-alive. This PR restores the handshake for generics without reverting #222 merge loading.

Root cause

Shape	Library .oxart?	Macro anchor ref?	Result
Concrete in lib (#72)	Yes	Yes	Works
Generics mono only in binary (#222)	Usually no	No (was skipped)	Works via merge
Generics mono in library, thin binary (#1365)	Yes	No (bug)	NoModules

Approach

Preserves the intentional design layers:

1. ModuleNotFound after I move a #[cuda_module] into a separate crate #72 archive handshake: keep referencing the package (or v2) artifact anchor from load_named.
2. Panic DriverError(500, "named symbol not found") when calling generic kernels from a separate crate #222 merge load: generic modules still call load_all_ptx_bundles_merged (no named-only regression).
3. No false undefined symbols: when the defining crate has merge markers but kernel_count == 0, the backend emits a strong anchor-only .oxlink stub (no fake empty PTX payload).

Macro (cuda_module_artifact_anchor_references)

• Remove the early return that skipped when every kernel is generic.
• Emit the same cfg-guarded black_box anchor refs for generic kernels via effective_cfg_attrs.
• Keep owner-filter / missing CARGO_PKG_* skips.

Backend (rustc-codegen-cuda)

• Extract cgus_require_ptx_bundle_merge so markers are scanned even without device code.
• When owner-selected, markers present, and there is no device codegen path (kernel_count == 0 / no device fns), emit a strong primary anchor-only object (v2 + weak legacy when owner filter is active).

Tests / example

• Macro unit: all-generic expansion has keep-alive + merge loader; cfg-gated generic gates the ref.
• oxide-artifacts: strong anchor-only object tests.
• New generic_mono_in_lib two-crate example matching the issue (lib mono + thin bin).
• Existing cuda_module_in_lib / cross_crate_embedded shapes remain the concrete and Panic DriverError(500, "named symbol not found") when calling generic kernels from a separate crate #222 regressions.

Orthogonal to #1367 / #1166 (finalizer expected-kernel inventory).

Test plan

• cargo test -p oxide-artifacts --features object --lib (24 passed, including new strong-anchor tests)
• scripts/sync-example-locks.sh --check
• scripts/check-example-smoketest-contract.sh
• cargo test -p cuda-macros --lib (blocked on this box: no CUDA 13 toolkit for cuda-bindings)
• cargo oxide run generic_mono_in_lib and --verify-bundles (no GPU / no toolkit here)
• cargo oxide run cuda_module_in_lib and cross_crate_embedded (same)

CI on this PR should cover the CUDA-backed lanes.

Reviewers

cc @nihalpasham @xavierforge (prior #72 / from-anchor / embedding owners; CODEOWNERS currently declares no path owners in-tree)

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.