bone_sse: pure Zig SSE/FMA reimplementation, zero game function calls

Replace bone_sse.zig with a complete pure Zig implementation compiled
with SSE4.1 + FMA + AVX. All 18 game function calls replaced:

- findInterpIdx (0x713D50): temporal-coherence keyframe search
- interpAnimKF (0x713EA0): CompQuat lerp for rotation keyframes
- extractByte (0x71AE90): byte keyframe extraction
- getInterpolatedFloat (0x71AF20): float track with direct blend read
- callFtol (0x40A2B0): @intFromFloat replaces x87 __ftol
- callVec3SqMag (0x4549F0): inline FMA dot product
- callGetIndexOffset/callSetShortValue (0x71AFF0/0x71B010): direct ri16
- matMul (0x74A7C0): V4 FMA matmul (broadcast + 3 @mulAdd per row)
- buildRotFn (0x74B6B5): inline quat→matrix
- rotateQuat (0x7BDDB0): quat→matrix then FMA matmul
- scaleMat (0x7BDCA0): inline scale from vec3 ptr
- applyTrans (0x7BDC40): inline FMA dot product translation

Only 2 game calls remain:
- 0x409AEF: one-time atexit init (boneKeyframeLoop)
- 0x7B5F60: IsParticleBufferEmpty (reads game particle state)

Child recursion calls transformMatrix4x4_SSE directly instead of
going through the hook at 0x714260.

Detour cleaned up: REF is baseline, SSE activates via ab_use_custom
toggle. Diagnostic/bisect/FPU-comparison scaffolding removed.
build.zig: bone_sse gets dedicated target with sse4_1+fma+avx features.

build.zig

@@ -63,11 +63,17 @@ pub fn build(b: *std.Build) void {
             .optimize = .ReleaseFast,
         }),
     });
+    const bone_sse_target = b.resolveTargetQuery(.{
+        .cpu_arch = .x86,
+        .os_tag = .windows,
+        .abi = .msvc,
+        .cpu_features_add = std.Target.x86.featureSet(&.{ .sse, .sse2, .sse3, .sse4_1, .fma, .avx }),
+    });
     const bone_sse_obj = b.addObject(.{
         .name = "bone_sse",
         .root_module = b.createModule(.{
             .root_source_file = b.path("src/transform44/bone_sse.zig"),
-            .target = target,
+            .target = bone_sse_target,
             .optimize = .ReleaseFast,
         }),
     });