fix python binding

[pca006132]

Fixes #947 (comment)

[pca006132]

Not entirely sure why the test is failing. May need a docker to test this.

[pca006132]

I guess warp_batch was broken... will see how to fix that.

[pca006132]

In fact, I am not sure if we should keep the warp_batch API. I think this kind of modifying array by reference is quite brittle with transfer of ownership. Returning a new array will always work, and I don't think that will add too much overhead.

[elalish]

Returning a new array sounds good to me.

[pca006132]

Summary:

1. No longer depends on unstable API. I found the solution to be surprisingly simple: just use the type caster for ndarray, it automatically performs conversion from ndarray to python handle (which was why we were trying to call ndarray_wrap).
2. Changed the warp_batch API to require returning an array. I feel that we were lucky the previous code works, because it is probably relying on numpy not changing the underlying buffer for some operations.
3. Added docstring override functionality, so we can customize some docstring. I think we may also want to add replacement function later, for example to replace different function names due to differences in naming convention.
4. Download 2.2.0 of nanobind, and set the version string if we download it using FetchContent (nanobind is unavailable in python search path, so we cannot get the version string automatically in this case).
5. Limit nanobind dependency to 2.2.0. I don't think we relied on unstable APIs now, but to avoid potential future build failure, it is better to specify the latest version that we support. As long as we have regular releases, we can still update nanobind version.