fix(BACKEND-GATE-ROCM-SGLANG): query shared memory through pinned HIP driver

Row: BACKEND-GATE-ROCM-SGLANG

Owner: current Strix benchmark campaign operator. Found during #3053 restoration. The reviewed preparation CLI at c8d019447 queries torch.cuda.get_device_properties(0).shared_memory_per_block, but the pinned AMD Torch 2.9.1 wheel does not expose that Python property. After repairing MIOpen, Strix job c13a24ee-d1ad-49a7-ba08-e96e814273c0 fails identity-dependencies with AttributeError before kernel compilation. Evidence: /mnt/nas_share/rc/strix-four-engine-3053.X94a3J/prepare-c8d019447-rebuilt-02/logs/identity-dependencies.log.

The pinned Triton AMD backend already returns max_shared_mem from HIP hipGetDeviceProperties: triton/backends/amd/driver.c lines 199-209; driver.py lines 161-177 and 700-705. Validate that executing path on Strix, then use the measured HIP value without changing the Torch/Triton pin, hard-coding 65536 as a measurement, or weakening the existing 64KiB identity gate. Add a regression that executes the production identity probe with Torch properties lacking the unsupported member. Preserve failure behavior for missing, invalid, or insufficient device limits. A committed design supplement precedes implementation; a fresh implementer and independent mutation reviewer are required. #3053 owns downstream preparation, model startup, and benchmarks.

Start with triton/backends/amd/driver.c lines 199-209 and driver.py lines 161-177 and 700-705, then trace the production identity probe used by the preparation CLI. Validate the HIP shared-memory path on Strix and add a regression where Torch properties lack the unsupported member. Done means valid HIP limits are used without changing pins or the 64KiB gate, while missing, invalid, or insufficient limits still fail.