Improvements to the sanitizer workflow for the Github CI

[oskooi]

1. MSAN is impractical for this project

MSAN (Memory Sanitizer) detects uninitialized memory reads, not memory leaks. The real reason it doesn't work (and has been commented out in build-san.yml since #3014) is that MSAN requires all linked code — to be compiled with -fsanitize=memory. The system packages (libfftw3, libhdf5, libgsl, libharminv, etc.) are not MSAN-instrumented, so every call into them produces false positives. Making MSAN work would require building all dependencies from source with MSAN, which is a significant effort. The recommendation is to remove the commented-out MSAN entry entirely rather than leaving it as dead config.

2. Dependency caching

Same as the main CI workflow changes in #3194 — the apt-get install step is fast since it uses pre-built packages, but caching would still help.

[oskooi]

Note: it is not necessary to change the argument runs-on from ubuntu-24.04 to ubuntu-latest.

This is because ubuntu-latest is an alias that Github periodically updates to point to a newer version, which can unexpectedly change the clang version, system libraries, and sanitizer runtime behavior — potentially introducing spurious failures unrelated to the Meep code. Pinning to ubuntu-24.04 gives reproducible results, which is especially important for sanitizer workflows where runtime behavior is sensitive to toolchain versions.

[oskooi]

As additional information, UBSAN (Undefined Behavior Sanitizer) detects C/C++ undefined behavior at runtime, such as:

• Signed integer overflow
• Null point dereferences
• Out-of-bounds array access
• Misaligned pointer access
• Division by zero
• Invalid type casts
• Shift operations past bit width

These are bugs where the C++ standard says "anything can happen" — the code may appear to work on one compiler/platform but crash or produce wrong results on another. UBSAN instruments the code at compile time and reports violations when the tests run.

[oskooi]

The sanitizer checks for ASAN and UBSAN (which were run manually) are passing.