Add conv3d backward tolerance overrides for CUDA (cherry-pick from #173188)

[srinivamd]

Summary

Cherry-pick three DecorateInfo tolerance overrides from upstream pytorch/pytorch PR #173188 ([ROCm][CI] Upgrade ROCm CI to 7.2 - 4/N, commit 8301e14b7003, merged 2026-02-22 by @jithunnair-amd).

These overrides relax float32 tolerance for nn.functional.conv3d backward on CUDA (atol=5e-5, rtol=5e-6) for:

• TestOperators.test_jvpvjp
• TestOperators.test_vjp
• TestCompositeCompliance.test_backward

Root Cause

TestCompositeComplianceCUDA.test_backward_nn_functional_conv3d_cuda_float32 flakily fails on ROCm (ROCM-25028) because:

1. MIOpen dilation workspace bug: A legacy guard in GetWorkSpaceSizeGEMM returns workspace=0 for dilation > 1, blocking preferred GEMM solvers (GemmBwdRest/GemmWrwUniversal) and forcing fallback to a less-accurate solver. Fix rocm-libraries#6507 was merged but reverted due to OOM on large diffusion workloads.

2. Missing tolerance override: The fallback solver produces gradient error of ~2.5e-5 — within the atol=5e-5 override on release/2.12 but exceeding the default atol=1e-5 on release/2.10.

The upstream tolerance override was added after the release/2.10 branch point (~Oct 2025) and is already present on release/2.11 and release/2.12. Intel XPU documented the identical pattern (pytorch/pytorch PRs pytorch#177069, pytorch#177848).

Test plan

• Verify PYTORCH_OPINFO_SAMPLE_INPUT_INDEX=10 PYTORCH_TEST_WITH_ROCM=1 python test/test_ops.py TestCompositeComplianceCUDA.test_backward_nn_functional_conv3d_cuda_float32 passes consistently on MI308/gfx94X
• Verify no regressions in test_ops.py -k TestCompositeComplianceCUDA test suite

Fixes: ROCM-25028

[rocm-repo-management-api]

Jenkins build for cdf49e3640ec27dfb26f0713528820b2aeb9e419 commit finished as FAILURE
Links: Pipeline Overview / Build artifacts / Test Results

[srinivamd]

PR Evaluation Summary

Code change: Correct and safe. The 3 DecorateInfo tolerance overrides are consistent with the existing test_vjpvmap override (same atol=5e-5, rtol=5e-6), properly scoped to device_type="cuda", and correctly placed in the decorators tuple.

Jenkins failure: Infrastructure issue, not code. The "This commit cannot be built" error is unrelated to this 12-line test decorator change — likely a Jenkinsfile branch config or Docker image pull issue. Recommend retriggering the build.

Two corrections needed in the PR description:

1. Upstream provenance: PR [ROCm][CI] Upgrade ROCm CI to 7.2 - 4/N pytorch/pytorch#173188 shows merged: false in the GitHub API because PyTorch's merge-bot uses a rebase workflow. The content did land on upstream main as commit 8301e14b7003 on 2026-02-22. Worth clarifying to avoid confusion during review.

2. Release branch claim is incorrect: The description states these overrides are "already present on release/2.11 and release/2.12" — they are not present on any ROCm/pytorch release branch. On upstream main, the conv3d opinfo was later refactored out of common_methods_invocations.py into test/functorch/test_ops.py, so these overrides were never back-ported to the ROCm release branches.