Review fixes for the GPU LogUp aux build (#762)

[MauroToscano]

Fixes from a review of #762. No behavior change to the proving path beyond freeing device memory earlier; the proof output is untouched.

VRAM (the substantive one)

The trace-domain snapshot retained by the R1 main LDE (GpuLdeBase.trace_dev + TraceTable.main_trace_dev) is consumed exactly once — by the LogUp aux build — but both Arcs lived to the end of the proof, holding a main-trace-sized device buffer (~3 GB on ethrex dims) through the aux-commit + DEEP/FRI VRAM peak, including for tables that never take the GPU aux path. Both Arcs are now dropped right after the aux-build pass, so the reuse win is kept and the resident window shrinks to exactly where it's needed.

Hardening

• cfg() in math-cuda/src/logup.rs now returns Err past the u32 grid limit instead of silently truncating the launch (leaving uninitialized output tails). Same pattern and rationale as batch_inverse_ext3_dev.
• num_out_cols == 0 in logup_aux_resident is a runtime Err instead of a debug_assert: in release it wrapped num_out - 1 and launched the assemble kernel with a bogus column count.
• Descriptor column indices are validated against the table width before any kernel launch. The kernels index main[col*num_rows + row] unchecked (they are never told num_cols), so a mis-authored table was a silent out-of-bounds device read where the CPU path panics cleanly.
• logup_aux_resident asserts num_rows >= 1 (it reads row_sum[(num_rows-1)*3..]).

Tests

• The term-column and resident-aux parity tests now canonicalize both sides before comparing (the canon3 pattern from the batch_inverse tests). The GPU pipeline computes the same field values through different op trees (tree-scan inverse, closed-form accumulate), so raw limbs can differ in the rare non-canonical band — a latent flake. The fingerprint test intentionally stays raw: that kernel mirrors the CPU evaluator op for op.
• Clippy findings in the cuda-gated test module fixed (that module is never linted by CI, which runs clippy without the cuda feature).

CI

New test-stark-cuda-lib job runs the CPU-runnable descriptor/CPU-mirror parity tests (cargo test --release -p stark --features cuda --lib logup_gpu) on a plain runner — without nvcc the kernels build as empty PTX stubs and the descriptor tests never touch the driver. Previously nothing in CI even compiled logup_gpu, since no job enables the cuda feature. (The GPU-only #[ignore] tests still need a manual run on the GPU box; note make test-cuda-integration targets -p lambda-vm-prover only, so it does not run the stark-lib GPU tests either.)

Docs

• Reattached three doc comments orphaned by items inserted beneath them (coset_lde_row_major_inner, u64_to_ext3_vec, logup_aux_resident) and updated the inner LDE return-tuple doc to the new 4-tuple.
• try_expand_..._keep_dev borrows the resident buffer (D2D copy) rather than consuming it; GPU_LOGUP_CALLS counts successes, not attempts.
• The disable_event_tracking comment now states the actual cross-stream invariant (producer host-syncs before its handle escapes) — this PR's trace_dev/ResidentAux handoffs made the old "we never share slices across streams" claim false.
• Deduplicated split_interactions (reuse the lookup.rs definition) and fixed the build_accumulated_column_from_terms name in logup.cu.

Validation

• cargo test -p stark --lib: 169 passed.
• cargo test -p stark --features cuda --lib logup_gpu: descriptor parity tests pass on a GPU-less machine (GPU tests stay ignored).
• cargo clippy clean on stark + math-cuda both with and without the cuda feature.
• Not run here (no GPU): the #[ignore] GPU parity tests and the cuda integration test — worth one run on the GPU box before merging.