Update CUTLASS and align build flags with newer headers

[ayushozha]

Summary

• update the CUTLASS submodule to the upstream typo fix for set_slice3x3
• compile tiny-cuda-nn as C++17/CUDA17 so the newer CUTLASS cute headers build cleanly
• enable the standard-conforming MSVC preprocessor for CUDA host compilation on Windows
• update the README compiler requirement to match the new baseline

Why

Issue #530 points out that the current bundled CUTLASS revision fails to compile because of the set_slice3x3 typo in matrix.h. The direct bump in PR #531 moves too far forward and trips the build because newer CUTLASS requires a newer language mode and a conforming MSVC preprocessor.

This change takes the smaller upstream CUTLASS fix commit and makes the project build settings match what that dependency now expects.

Validation

• configured from scratch with CUDA 13.2 and Visual Studio 2022
• cmake -S . -B build-cutlass-check -G Ninja -DTCNN_BUILD_TESTS=1
• cmake --build build-cutlass-check --target test_grid
• test_grid.exe
• ctest --output-on-failure -V -R test_grid

Fixes #530.

[Tom94]

Much appreciated!