Fix 32-bit fixed-point multiply to use a correct 64-bit intermediate

[RichardHoekstra]

Problem

The 32-bit path of FixedMul (Q16.16) is wrong for any operand with a nonzero integer part. It splits each operand into hi/lo halves, but:

result_hi = a_hi * b_hi                  ' missing the << 16
result_lo = (a_lo * b_lo) SHR 16
result = result_hi + result_lo + ((a_hi * b_lo) SHR 16) + ((a_lo * b_hi) SHR 16)

The correct expansion of (a*b) >> 16 is a_hi·b_hi·2^16 + (a_hi·b_lo + a_lo·b_hi) + (a_lo·b_lo)>>16. Two terms are misplaced: result_hi is missing its << 16, and the cross terms are shifted right instead of kept at weight. Only the pure-fraction term lands correctly, so the result is right only when both operands are < 1.0.

Measured in the real DOS runtime (before the fix):

product	expected (Q16.16)	fallback returned
0.25 × 0.25	4096	4096 ✓
2.0 × 0.5	65536	1
3.0 × 3.0	589824	9
1.5 × 2.0	196608	3
10.0 × 0.1	65536	1

Where it bites

• __ASM_AVAILABLE__ is never defined, so FixedMulAsm's real inline-asm block never compiles — its simulated #ELSE body is the same buggy formula as FixedMultiplyFallback. So HasAssemblySupport() doesn't gate it: both 32-bit branches are wrong.
• It's masked on 64-bit builds, which take the correct #IFDEF __FB_64BIT__ branch — which is why it isn't visible in development.
• On 32-bit DOS it affects the lab / reference transformer path (attention scaling, LayerNorm, GELU, softmax) and InitQuantizationTables.
• The shipped preview is unaffected: main_prod → real_gpt.bas uses its own correct TinyGPTFixedMul ((CLNGINT(a)*CLNGINT(b)) \ 4096).

Fix

Both 32-bit branches now use a 64-bit LONGINT intermediate (emulated on 32-bit targets), matching the existing __FB_64BIT__ branch and TinyGPTFixedMul:

DIM result AS LONGINT
result = CLNGINT(a) * CLNGINT(b)
result = result SHR FIXED_POINT_SHIFT
RETURN CLNG(result)

Verification

Re-run of the same DOS repro against the patched functions — every case now matches a 64-bit reference exactly, including negatives:

product	ref	fallback	asm-sim
0.25 × 0.25	4096	4096	4096
2.0 × 0.5	65536	65536	65536
3.0 × 3.0	589824	589824	589824
1.5 × 2.0	196608	196608	196608
10.0 × 0.1	65540	65540	65540
−2.0 × 3.0	−393216	−393216	−393216

• Cross-compiled with DJGPP FreeBASIC, run under DPMI (dosbox-x).
• src/main_prod.bas cross-compiles to DOS clean with this change.
• python3 -m unittest discover tests → 108 passed; git diff --check clean; verify_workspace_tracking.py passes.
• Diff is one file, two functions in src/asm_optimizations.bas (+16 / −31).