Refactor: Lazily load or eliminate heavy dependencies (scipy, numba, h5py, astropy, cosmopy) to fix import overhead

[kayhangultekin]

Description

This PR resolves the critical package initialization bottleneck by completely decoupling heavy scientific packages (scipy, numba, h5py, astropy, and cosmopy) from the global top-level import sequence.

Key Changes

• Lazy-Loaded Cosmopy Integration: Deferred the initialization of cosmopy by encapsulating it in a lazy caching function. This completely eliminates the significant overhead of setting up the default cosmology framework on package boot.
• Eliminated Astropy from Core Constants: Rewrote holodeck/constants.py to use hardcoded physical constant values rather than pulling from astropy.constants. Because constants.py is loaded directly in __init__.py, this completely immunizes the base package initialization from the astropy constants registry overhead.
• Hardcoded 1-Sigma Constant: Replaced a dynamic scipy.stats.norm.cdf(1) call with its exact float64 constant (0.8413447460685429), eliminating a global scipy load in utils.py.
• Localized SciPy & h5py: Shifted scipy.interpolate (for PchipInterpolator), scipy.special (for jn), and h5py imports directly into the specific function scopes where they are executed.
• Lazy Numba JIT Decorator: Implemented a lightweight proxy decorator (_LazyNJIT) around _gw_ecc_func. This completely bypasses loading the heavy LLVM compiler machinery at boot time, delaying the import until the function is actually called during execution.
• LaTeX Syntax Fix: Prepend r to the docstring on line 1073 to resolve a Python 3.12 SyntaxWarning regarding an invalid escape sequence (\l).

Performance Impact

Using time python -c "import holodeck", package startup overhead dropped by over 95% on subsequent runs (2600% if Trump is doing the math), dropping total wall-clock time from ~2.6 seconds down to around 0.1 seconds.

Side-by-Side Comparison:

⟩ git switch dev
Switched to branch 'dev'
Your branch is up to date with 'origin/dev'.
⟩ for i in $(seq 10 ) ; do time python -c "import holodeck" ; done
/Users/kayhan/Documents/projects/nanograv/holodeck/holodeck/utils.py:1058: SyntaxWarning: invalid escape sequence '\l'
  You may need to convert, e.g. $dN/dx = \frac{dN}{d \ln x} \frac{1}{x}$.
python -c "import holodeck"  1.37s user 0.31s system 52% cpu 3.186 total
python -c "import holodeck"  1.32s user 0.23s system 59% cpu 2.594 total
python -c "import holodeck"  1.30s user 0.24s system 59% cpu 2.577 total
python -c "import holodeck"  1.31s user 0.24s system 60% cpu 2.582 total
python -c "import holodeck"  1.31s user 0.24s system 59% cpu 2.584 total
python -c "import holodeck"  1.30s user 0.24s system 59% cpu 2.580 total
python -c "import holodeck"  1.31s user 0.24s system 59% cpu 2.576 total
python -c "import holodeck"  1.33s user 0.24s system 60% cpu 2.608 total
python -c "import holodeck"  1.33s user 0.24s system 60% cpu 2.618 total
python -c "import holodeck"  1.31s user 0.24s system 60% cpu 2.577 total

⟩ git switch dev-kg-functioncaching
Switched to branch 'dev-kg-functioncaching'
Your branch is up to date with 'origin/dev-kg-functioncaching'.
⟩ for i in $(seq 10 ) ; do time python -c "import holodeck" ; done
python -c "import holodeck"  0.07s user 0.03s system 57% cpu 0.181 total
python -c "import holodeck"  0.07s user 0.02s system 95% cpu 0.097 total
python -c "import holodeck"  0.07s user 0.02s system 92% cpu 0.096 total
python -c "import holodeck"  0.07s user 0.02s system 94% cpu 0.093 total
python -c "import holodeck"  0.07s user 0.02s system 89% cpu 0.100 total
python -c "import holodeck"  0.07s user 0.02s system 90% cpu 0.101 total
python -c "import holodeck"  0.07s user 0.02s system 91% cpu 0.108 total
python -c "import holodeck"  0.08s user 0.03s system 86% cpu 0.117 total
python -c "import holodeck"  0.07s user 0.02s system 91% cpu 0.101 total
python -c "import holodeck"  0.07s user 0.02s system 87% cpu 0.106 total

[astrolamb]

Hey @kayhangultekin this looks great to me. I'm also seeing a similar speed-up on import. There's just some clean-up required for this PR.

I think we can also drop the requirement for python 3.9, which is the only version currently failing tests

[astrolamb]

@kayhangultekin it looks like we have an assertion issue with the ubuntu tests... macos is ok though