Print full-precision mismatch values in CPU and GPU test helpers

[IvanaGyro]

Problem

AreNearlyEqStorage prints mismatched values through std::cout using the default stream precision. For floating-point differences below roughly six significant digits, the comparison can fail while the diagnostic prints two values that look identical.

This was noted in the review discussion on #1148:
#1148 (comment)

The same implementation exists in both test helper paths:

• tests/test_tools.cpp for CPU tests
• tests/gpu/gpu_test_tools.cpp for GPU tests

Proposed change

Use sufficient precision when printing floating-point and complex mismatch values in both helpers. For double-precision values, this can be std::setprecision(17) or, preferably, std::numeric_limits<double>::max_digits10.

The implementation should avoid leaking modified formatting state through the global std::cout stream, for example by restoring the previous precision or formatting through a local stream.

Acceptance criteria

• CPU and GPU AreNearlyEqStorage diagnostics show distinguishable values when a mismatch occurs beyond the default six-digit precision.
• Complex values receive equivalent full-precision diagnostics.
• Integer and boolean diagnostics remain readable and unchanged in meaning.
• Printing a mismatch does not permanently change the formatting state of std::cout.

— Issue opened by GPT-5.6 Thinking on behalf of @IvanaGyro