ENH: Add data loader class, install at root and data modules

[effigies]

Fixes current test failures from using a deprecated importlib.resources API.

Expanded on an ExitStack-based load function that @mgxd used in one of these projects into a class that provides flexible access to package data with readable, unzipped, and cached access methods.

We have the following loaders:

• niworkflows.load_resource - Generic, can get anything from anywhere
• niworkflows.data.load - Primary access point for package data
• niworkflows.tests.data.load_test_data - Convenience for tests
• niworkflows.interfaces.tests.data.load_test_data - Convenience for interface tests

I would like to figure out how to bundle the library in a zip to actually test that this works when installed that way, but for this PR the goal is to make it easy to think about and select whether we need the resource to be a Path or just Traversable.

This also finally purges pkg_resources, which we should avoid because we have no runtime dependency on setuptools, and the API is discouraged in third-party packages.

[effigies]

Tests passing. Anybody have time for a review?

[mgxd]

This is cool! If I'm understanding it right, this can be used as a dataloader across the *preps ecosystem?

[mgxd]

Since it seems like this step is often, if not always, followed with a read_text() call, wdyt about adding another method (read_text?) to serve as some syntactic sugar?

[effigies]

I like the explicitness of loader.how_to_load(what_to_load).do_something().

[effigies]

If I'm understanding it right, this can be used as a dataloader across the *preps ecosystem?

Yes. I would probably just vendor it in the local data/__init__.py for a few versions, rather than depend on it in niworkflows.

I think it's a handy util but not sure it's stable enough to really want to deal with it as a dependency. I did an initial pass in BIDS, and then while working on this refactor found I wanted to explicitly avoid caching.

[oesteban]

looks great!

[oesteban]

Yes. I would probably just vendor it in the local data/__init__.py for a few versions, rather than depend on it in niworkflows.

Just to confirm - "vendor it in" means copying the infrastructure across packages (for a few versions)?

[effigies]

Yes