[ROCm] Skip TF32 dot algorithm tests on Navi/RDNA GPUs

[srinivamd]

Summary

Cherry-pick the Navi/RDNA TF32 skip guard from rocm-jaxlib-v0.8.2 to rocm-jaxlib-v0.9.1.

• Skip TF32_TF32_F32 and TF32_TF32_F32_X3 dot algorithm tests on Radeon/RDNA GPUs
• TF32 is only supported on CDNA GPUs (MI100+/gfx9) via MFMA instructions; RDNA GPUs lack this hardware
• The test was being skipped on v0.8.2 but fails on v0.9.1 due to upstream jax-ml/jax#34534 changing the skip guard to CUDA-only

Root Cause

Upstream PR jax-ml/jax#34534 (merged 2026-02-11 by magaonka-amd) changed the TF32 skip logic from:

if not jtu.is_cuda_compute_capability_at_least("8.0"):

to:

if jtu.test_device_matches(["cuda"]) and not jtu.is_cuda_compute_capability_at_least("8.0"):

This correctly enables TF32 tests on ROCm CDNA GPUs (MI100+) but also unblocks them on RDNA GPUs (Navi/Radeon) where XLA's gfx9_mi100_or_later() gate correctly returns UNIMPLEMENTED.

The rocm-jaxlib-v0.8.2 branch had a supplementary skip block (via ROCm/jax#635) that checked "Radeon" in device_kind to skip TF32 on RDNA GPUs. This was never upstreamed or cherry-picked to v0.9.1.

Test plan

• Run tests/lax_test.py::LaxTest::testDotAlgorithm13 on Navi4x/gfx1201 — should now be SKIPPED instead of FAILED
• Run full testDotAlgorithm suite on MI300X/gfx942 — TF32 tests should still PASS
• Run full testDotAlgorithm suite on Navi4x — non-TF32 tests should still pass, TF32 variants skipped

Fixes: ROCM-25363