[JonathanLichtnerAMD]

Motivation

Fixes #8065. Full/GPU_SoftMarginLossForward_* with reduction_mode MEAN (2) or SUM (1)
failed in the -L full CI sweep on gfx94X-dcgpu (MI300X) with hipError 709
("context is destroyed"), blocking the rocm-systems → TheRock bump PRs.

Root cause: the reduction invoker in
projects/miopen/src/solver/softmarginloss/forward_softmarginloss.cpp passed a
tensor_view_t<5> (80 B) as the last argument to the final ReduceSum kernel,
whose signature takes tensor_view_t<1> (16 B). The 64 B overflow corrupted
the kernarg buffer; the kernel then used garbage stride/offset fields for
address computation, and the resulting wild GPU stores corrupted HSA pool
tracking. The deferred hipError 709 was raised later by the HIPErrorHandler
gtest listener at test teardown.

Risk Assessment

Low. Five-line change confined to a single MIOpen solver
(SoftMarginLossForward). The change replaces an incorrectly-sized
tensor_view_t argument with the correctly-sized one — matching the kernel's
declared signature — and removes manual hipEventDestroy calls that duplicate
the work HipEventPtr's RAII deleter already performs (matching the canonical
HipEventProfiler pattern in src/hipoc/hipoc_kernel.cpp). No public API,
ABI, header, schema, or build-system changes. No algorithmic changes. The
reduction_mode = NONE path is untouched.

Technical Details

1. Build a tensor_view_t<1> from the output descriptor and pass it to the
   final ReduceSum kernel instead of the 5D o_tv used by the per-element
   phase.
2. Remove the two (void)hipEventDestroy(...) calls in the if(profiling)
   cleanup block — HipEventPtr (a MIOPEN_MANAGE_PTR(hipEvent_t, hipEventDestroy) alias) destroys both events on scope exit.

Test Plan

Manually reproduced and verified on gfx942 (MI300X) under AddressSanitizer
(-fsanitize=address -shared-libasan, HSA_XNACK=1) using the MIOpen gtest
binary. CI should handle the rest.

Test Result

• Manual ASAN tests passed.
• All 90 Full/GPU_SoftMarginLossForward_FP32.* tests passed.

CI will exercise the FP16 / BFP16 variants automatically.

Submission Checklist

• Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.