Relax NaN comparison in compression tests (#2388)

* Relax NaN comparison in compression tests

`compareExact(float)` currently requires bitwise equality, even for
NaN, but C vs C++ EXR can generate NaNs that are not bit-wise
identical, leading to reports of failures on various architectures.

This fixes the failures noted in #1281, #1460, #1628 in
the `test*Compression` tests, of the form:

    F float at 59, 0 not equal: C++ loaded C 0x7fc4e364 (nan) vs C loaded C 7f84e364 (nan)

and in `testOptimizedInterleavePatterns`:

    error reading back channel B pixel 21,-76 got -nan expected -nan

This change takes the conservative approach and relaxes the bitwise
comparison only for the architectures that are known to have NaN
differences.

This also changes the method of bitwise comparison of float
values. The C++ standard does not guarantee that writing one field of
a union and reading another is well-defined, and we've seen cases of
it failing with Intel's oneAPI. This changes the comparison to use
`memcpy` to a `uint32_t` instead of a union.

Made with Cursor
Signed-off-by: Cary Phillips <cary@ilm.com>

* remove extraneous bothNan

Signed-off-by: Cary Phillips <cary@ilm.com>

* Use  test on all architectures

Signed-off-by: Cary Phillips <cary@ilm.com>

---------

Signed-off-by: Cary Phillips <cary@ilm.com>

Author: cary-ilm

src/test/OpenEXRCoreTest/compression.cpp

@@ -519,22 +519,24 @@ struct pixels
         const char* tagb,
         const char* chan)
     {
-        union
-        {
-            uint32_t iv;
-            float    fv;
-        } a, b;
-        a.fv = af;
-        b.fv = bf;
-        if (a.iv != b.iv)
+        // On some architectures (Windows 32‑bit x86, i686 with x87 extended
+        // precision, aarch64, ppc64le, arm7), the C and C++ codepaths may
+        // generate NaN values that are not bitwise identical. Accept these
+        // special cases as equal.
+        const bool bothNaN = std::isnan(af) && std::isnan(bf);
+        if (bothNaN)
+            return;
+
+        uint32_t a,b; memcpy(&a,&af,sizeof a); memcpy(&b,&bf,sizeof b);
+        if (a != b)
         {
             std::cout << chan << " float at " << x << ", " << y
-                      << " not equal: " << taga << " 0x" << std::hex << a.iv
+                      << " not equal: " << taga << " 0x" << std::hex << a
                       << std::dec << " (" << af << ") vs " << tagb << " "
-                      << std::hex << b.iv << std::dec << " (" << bf << ")"
+                      << std::hex << b << std::dec << " (" << bf << ")"
                       << std::endl;
         }
-        EXRCORE_TEST (a.iv == b.iv);
+        EXRCORE_TEST (a == b);
     }
     void
     compareExact (const pixels& o, const char* otag, const char* selftag) const