[rocm-libraries] ROCm/rocm-libraries#8762 (commit fe88750)

[CK] Fix flaky test `test_batched_gemm_b_scale_wmma`

Loosen up the tolerance and add better logging in case of failure

TLDR: `profile_batched_gemm_b_scale_impl.hpp` (lines 390-392) uses `rtol
= atol = 1e-1`, while the sibling non-batched B-scale profiler uses
`2e-2` in `profile_gemm_b_scale_impl.hpp` (lines 361-364). `KBatch > 1`
adds an F16 atomic accumulation step, so the batched path is stricter on
the noisier arithmetic path. Loosening up the tolerances should fix the
flaky test. I've also added better loging in case of failures.

## Motivation

`test_batched_gemm_b_scale_wmma` can fail on gfx1201 when `KBatch > 1`
because that path uses split-K F16 atomic accumulation. The existing
batched B-scale tolerance was stricter than the non-batched B-scale
profiler, even though the batched split-K path has additional rounding
noise from native F16 atomics.

This PR aligns the batched tolerance with the comparable non-batched
path and improves failure diagnostics so future mismatches identify the
exact shape, KBatch value, instance, strides, and operator.

## Technical Details

Updated
[projects/composablekernel/profiler/include/profiler/profile_batched_gemm_b_scale_impl.hpp](projects/composablekernel/profiler/include/profiler/profile_batched_gemm_b_scale_impl.hpp)
to:

- Change non-FP8 verification tolerance from `1e-2` to `2e-2`.
- Match the tolerance already used by the non-batched B-scale profiler.
- Build a detailed failure message with:
  - instance index
  - `kbatch`
  - `M`, `N`, `K`, and `BatchSize`
  - `StrideA`, `StrideB`, and `StrideC`
  - selected operator name
- Move operator name construction earlier so it is available for
diagnostics.

## Test Plan

Configured the CK gfx1201 build with device instances enabled and XDL
disabled to isolate WMMA coverage.

Built and ran the focused batched GEMM B-scale WMMA test target:

```bash
ninja -C projects/composablekernel/build-gfx950-gfx1201 test_batched_gemm_b_scale_wmma
```

Ran the focused `MidLargeM` case across WMMA instances and then ran the
full WMMA-only gtest target.

## Test Result

The WMMA-only gfx1201 validation passed.

- `MidLargeM` passed across 21 WMMA instances.
- Full `test_batched_gemm_b_scale_wmma` passed all 3 gtests.
- No incorrect-result lines were reported.

## Submission Checklist

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

Signed-off-by: Chris Tsiaousis <chris.tsiaousis@streamhpc.com>