Importing OCELOT is too slow

[st-walker]

We need to speed it up. It takes nearly 3 seconds to import ocelot. This is not acceptable.

The original sin is that the top level init.py imports basically everything. I guess we can't change that as everyone's code will then break.

Slow:

• importing some parts of scipy
• precompiling with numba in math_op.py and anywhere else, we should always compile lazily.
• importing pandas is slow
• h5py is not that slow but it should be optional.

Solutions:

• lazy compile all numba stuff
• move scipy imports into function bodies
• move to polars from pandas
• rewrite __init__.py to not import the entire package.
• h5py should be optional

with some the above stuff I got the import time down to 0.7s.

If we wrap the scipy imports then it would be faster still. Without moving all the scipy imports we can at least get down to 900ms import time or so.

Before, 2.7s:

After 0.7s:

[sergey-tomin]

implemented lazy import. now recommend way to import ocelot is 'import ocelot as ocl'. see 0729fe3