feat: batchgen_kernels versioning, ccache, JIT dev mode, and multi-arch support

[Andrewxu313]

Description

Add proper version management, incremental build support, and multi-architecture infrastructure to batchgen_kernels. Previously, the version was hardcoded as "0.1.0" in three places with no tagging, no compatibility checks, and full recompilation required for any change.

Motivation

1. No version tracking: Users had no way to know which kernel version was installed. Version mismatches between batchgen and batchgen_kernels caused silent failures.
2. Slow iteration: Adding or modifying a single .cu file required recompiling all 37 CUDA files (~10-20 min). No caching.
3. No multi-arch story: SM100 (Blackwell) kernels need entirely different source code (tcgen05 vs WGMMA), but the build system had no arch-awareness beyond compile flags.

Changes

1. Version infrastructure (single source of truth)

• New _version.py: __version__ = "0.2.0", version_info = (0, 2, 0), arch-aware __version_full__
• setup.py and pyproject.toml read version dynamically — one place to bump
• Git tag convention: batchgen-kernels-v{X.Y.Z}+sm90a

2. Runtime compatibility check

• New batchgen/kernel_compat.py: checks batchgen_kernels.version_info >= MIN_KERNELS_VERSION at import batchgen
• Clear error message with rebuild instructions if outdated

3. ccache/sccache integration

• setup.py auto-detects ccache or sccache and configures CC/CXX wrappers
• Unchanged files are cache hits (~instant). Only modified .cu files recompile.

4. JIT dev mode

• BATCHGEN_KERNELS_DEV=1 enables JIT fallback in load_extension()
• New _jit_registry.py mirrors all 22 CUDAExtension configs for JIT compilation
• Developers can add a new .cu file and test immediately without rebuilding the wheel

5. Multi-architecture build gating

• BUILD_ARCH env var: "sm90a" (default), "sm100", "all"
• SM90a and SM80+ extensions split into separate lists, gated by BUILD_ARCH
• Wheel naming: batchgen_kernels-0.2.0+sm90a-*.whl (PEP 440 local version)

6. External user support

• CUDA_HOME auto-detection from nvcc path
• Helpful error if CUDA toolkit not found

File	Change
batchgen_kernels/_version.py	New — single source of truth for version
batchgen_kernels/_jit_registry.py	New — JIT compilation config for all extensions
batchgen_kernels/__init__.py	Version import, load_extension() with JIT fallback
batchgen_kernels/setup.py	Dynamic version, ccache, CUDA_HOME, BUILD_ARCH gating
batchgen_kernels/pyproject.toml	Dynamic version via setuptools
batchgen/kernel_compat.py	New — runtime version compatibility check
batchgen/__init__.py	Add compat check at import

Type of Change

• Bug fix
• New feature
• Breaking change
• Code cleanup
• Documentation update

Checklist

• I have read the CONTRIBUTION guide.
• I have updated the tests (if applicable).
• I have updated the documentation (if applicable).