fix(kernel): enable disk row stores on windows

[wenlive]

Related to #384; complements #575.

Windows source builds omit _row_store, and the implementation depends on POSIX file I/O. This prevents using disk-backed PLE tables on Windows.

Implement the existing platform seams with Unicode file handles, overlapped positioned reads, aligned allocation and CUDA driver symbol loading. This lets fixed-width table rows remain on disk and enter the existing bounded staging buffers. The PLE hash and Linux reader algorithms retain their current behavior.

Extend the row-store tests with Unicode paths, offsets above 4 GiB, and recovery after a short read. Give the graph test's second store its own checkpoint file so a Windows writer does not overwrite a still-open file.

Validation: Windows 11 build 26200, RTX 5090 D v2 24 GB (driver 610.88), Ryzen 9 9900X, 64 GB RAM, MSVC 19.44, CUDA 13.0, Python 3.12.13, Torch 2.11.0+cu130. Base: 74a387777708b9650a027494bcb37d2ab9cf2443.

python build_row_store_validation.py --root . --with-pinned
$env:PYTHONPATH = 'python'
python -m pytest tests/kernels/test_row_store.py tests/models/qwen4_exp/test_ple_disk.py -q

11 passed, including GPU-oracle and CUDA-graph checks. Independently rebuilt against latest-main's Torch 2.14.1 CPU dependency: 8 passed, 3 GPU tests skipped (commands in the linked validation records). The baseline row-store source fails to compile on dlfcn.h with both versions.

The harness (source and test logs) selects the declared row-store extension and builds the pinned-memory prerequisite with Windows compiler/library settings. Synthetic byte tables and toy PLE checkpoints were used. Full-model serving, Linux hardware and Torch 2.14 CUDA were not validated; the 2.11 CUDA run does not verify main's complete dependency contract.

Prepared with AI assistance. The contributor has reviewed and understands the change; the reported tests were executed through development tooling.