[hipDNN] Improve sample failure output with tensor diffs

[Bingtagui404]

Description

When a CPU validation failure occurs in the hipDNN samples, the output only tells you which tensor failed — not how it failed:

CPU reference validation:
  y: failed

This PR adds a reusable tensor diff utility and wires it into all sample validation paths so failures automatically print detailed diagnostics:

CPU reference validation:
  y: failed
  Tensor diff for "y":
    Total elements: 65536
    Mismatched:     42 (0.06%)
    Max abs diff:   1.234567e-03 at [0, 2, 14, 7]
    Mean abs diff:  3.456789e-04
    Worst mismatches:
      [0, 2, 14, 7]: ref=0.543210, impl=0.544444, diff=1.234567e-03
      ...

Changes

• New: test_sdk/.../utilities/TensorDiff.hpp — header-only utility providing:
  • computeTensorDiff<T>() — element-wise comparison with summary statistics
  • printTensorDiffSummary() — formatted output
  • validateAndReport<T>() — drop-in replacement combining allClose() + status print + diff on failure
• Modified: All 14 sample .cpp files to use validateAndReport<T>() instead of manual allClose() + cout pattern

Safety

• Shape/element-count mismatches are detected before element-wise comparison to prevent out-of-bounds access
• computeTensorDiff runs single-threaded to avoid data races in summary accumulation
• maxMismatches == 0 is handled as "summary only" mode

Fixes #5547