assemble: fix broadcast ratio calculation for sub-64-bit vector operands

[agourakis82]

Fixes #263.

Summary

In asm/assemble.c, get_broadcast_num() previously calculated the number of broadcast elements with:

brcast_num = ((opsize / (BITS64 >> SIZE_SHIFT)) * (BITS64 / brsize))
    >> (opsize > (BITS64 >> SIZE_SHIFT));

When the operand size is smaller than BITS64 (such as xmmrm32 in VCVTPH2PD or VCVTPH2QQ with xmm destination, where opsize is BITS32 which equals 4, and BITS64 >> SIZE_SHIFT equals 8), the integer division opsize / 8 evaluated to 0.

As a consequence, brcast_num became 0, causing valid {1to2} broadcast instructions like:

vcvtph2pd xmm0, [rcx]{1to2}

to fail with:

error: mismatch in the number of broadcasting elements

Fix

Replace the calculation with an exact bit-size lookup table indexed by ilog2_32(opsize), accurately computing the broadcast ratio for all operand sizes from BITS32 through BITS512.

Verified assembling vcvtph2pd xmm0, [rcx]{1to2}, vcvtph2pd ymm0, [rcx]{1to4}, and vcvtph2pd zmm0, [rcx]{1to8}.

[agourakis82]

I rebuilt NASM locally from this PR head and ran the targeted AVX512-FP16 broadcast-related instruction tests.

HEAD=d247c04c3ad33a861ef22077cf1401cd30729d5e
./nasm -v
NASM version 3.02 compiled on Sep 29 2026

sh autogen.sh
sh configure
make -j4

python3 ./tools/travis/nasm-t.py -d ./travis/insns/vcvtph2pd --nasm ./nasm run --stop=n
# vcvtph2pd.1 .. vcvtph2pd.5 PASS

python3 ./tools/travis/nasm-t.py -d ./travis/insns/vcvtph2qq --nasm ./nasm run --stop=n
# vcvtph2qq.1 .. vcvtph2qq.5 PASS

python3 ./tools/travis/nasm-t.py -d ./travis/insns/vcvttph2qq --nasm ./nasm run --stop=n
# vcvttph2qq.1 .. vcvttph2qq.5 PASS

git diff --check origin/master...HEAD
# clean

I did not run the full Travis matrix locally, but the directly related conversion/broadcast instruction fixtures pass with the rebuilt assembler.