Fix float16_to_float(vint4) decoding every negative half as NaN on F16C builds

[lenamonj]

On an F16C build, float16_to_float(vint4) narrows the four lanes with _mm_packs_epi32, which saturates signed. A lane holding a half code with the sign bit set is a value above 0x7FFF, so it saturates to 0x7FFF and _mm_cvtph_ps decodes it as NaN. Every negative half decodes as NaN on AVX2 builds, while the SSE4.1, SSE2 and no-intrinsics backends decode it correctly.

The lanes hold zero-extended 16-bit codes, so _mm_packus_epi32 keeps all of them.

Test: SuiteVfloat4.float16_to_float now includes negative halves. test-unit-avx2 fails on main (all four lanes NaN) and passes with the change; the other unit binaries pass both ways. The x64 CI recipe was run locally: release and debug builds, ctest, functional and image tests for all four encoders, example build.

[solidpixel]

Is this behavior actually reachable? Do you have any cases that trigger astcenc to actually try to convert a negative value into fp16?

The only case where I think its possible gets clamped to zero before reaching the fp16 path.

[lenamonj]

Yes, through the input side rather than the encoder's float-to-fp16 path: load_texel_f16 in astcenc_image.cpp decodes half-float input texels with float16_to_float(vint4), so a half-float image with negative components reaches it as raw codes with the sign bit set. With no preprocessing the later clamp hides the difference, which matches what you describe. -pp-normalize and -pp-premultiply use the decoded values before that clamp, so a half-float normal map with negative components compresses differently on AVX2.

Reproducer: a 16x16 R16G16B16A16_FLOAT DDS whose texels cycle -1.0, 0.5, -0.5, 1.0, compressed with -cH in.dds out.astc 4x4 -medium -pp-normalize. On main, astcenc-avx2 gives 9e434c47..., astcenc-sse4.1 and astcenc-none give 11f96a63...; with this change all three give 11f96a63.... Same with -pp-premultiply (619b74c7... versus 0a8afc84...); identical everywhere with no preprocessing.

import struct
hdr = struct.pack('<7I', 124, 0x1007, 16, 16, 0, 0, 0) + b'\x00' * 44
hdr += struct.pack('<8I', 32, 0x4, 0x30315844, 0, 0, 0, 0, 0) + struct.pack('<5I', 0x1000, 0, 0, 0, 0)
dx10 = struct.pack('<5I', 10, 3, 0, 1, 0)
pattern = [0xBC00, 0x3800, 0xB800, 0x3C00]
body = b''.join(struct.pack('<H', pattern[(x + y + c) % 4]) for y in range(16) for x in range(16) for c in range(4))
open('f16neg.dds', 'wb').write(b'DDS ' + hdr + dx10 + body)