Support structured numpy dtypes in NDBuffer.copy_from_numpy()

[wlewNV]

Issue #636

Previously, passing ANY structured numpy dtype to NDBuffer.copy_from_numpy() would fail because nanobind's numpy binding doesn't accept compound dtypes.

This change does the following:

• copy_from_numpy() now uses BufferCursor for structured dtypes, which handles platform-specific alignment
• Simple dtypes (e.g., float32, int32) use a direct memory copy
• from_numpy() raises a helpful error for structured dtypes without explicit Slang type

[ccummingsNV]

This one should maybe wait until after the tensor refactor, because:

• We're getting rid of NDBuffer altogether (only Tensor will exist)
• We need to get rid of all Python side implemenation of the tensor type eventually, so we'll need to work out how to achieve the above from the native side (or at least via some hook)