fix(simd): build NEON lane-mask weights without brace init on MSVC

[adsharma]

Problem

Recent SIMD filter kernels (70f8828c5, d187dd03c) break Windows ARM64 builds. MSVC fails compiling src/function/simd_filter_neon.cpp:

simd_filter_neon.cpp(226): error C2078: too many initializers
simd_filter_neon.cpp(231): error C2078: too many initializers
simd_filter_neon.cpp(234): error C2078: too many initializers

Root cause

MSVC ARM64 maps the NEON vector types (uint8x8_t, uint16x8_t, uint32x4_t, ...) to the intrinsic __n64/__n128 types, which cannot be initialized with scalar brace lists (same known issue as xtensor-stack/xsimd#611). The lane-mask weights {1, 2, 4, 8, ...} in toLaneMask were constructed with = {...} initializers.

Fix

Build the weights from bit patterns using vcreate_u8 / vcreate_u16 / vcreate_u32 + vcombine, which is portable across GCC, Clang, and MSVC:

• uint8x8_t: vcreate_u8(0x8040201008040201ULL) → {1,2,4,8,16,32,64,128}
• uint16x8_t: vcombine_u16(vcreate_u16(...), vcreate_u16(...)) → {1,...,128}
• uint32x4_t: vcombine_u32(vcreate_u32(...), vcreate_u32(...)) → {1,2,4,8}

Verification

• Lane-mask extraction verified against scalar reference on ARM64 (all single-lane, full, and mixed patterns).
• SIMDFilterTest.* (4 tests) pass locally on ARM64.
• ci-workflow.yml with platform=windows — Build and Test steps pass on the Windows ARM64 runner.
• clang-format-18 clean.