Speed up GLTF loading.

[eliphatfs]

In my profile, more than 80% time is spent in the loop of strided accessor construction when loading GLTF/GLB files.

I thus applied np.lib.stride_tricks.as_strided to speed up the process.
I tested it on multiple glb files and the results are the same as the original logic.

This implementation also works for stride = 0, repeating the first element for count times. But glTF 2.0 specification 5.11.4 says it should be positive, so I added an assertion.

[mikedh]

Looks good!! The docs for as_strided do encourage "extreme caution" but it does look like you're checking all the things they warn about (copy the result, make sure data is long enough, etc). And you're right it's a nice speedup on corpus (although I'd also like to check that it loads the same number of meshes :)

This PR:
- 3.289 load_gltf  trimesh/exchange/gltf.py:268
- 3.054 load_glb  trimesh/exchange/gltf.py:343

Current `main`:
- 5.022 load_gltf  trimesh/exchange/gltf.py:268
- 3.616 load_glb  trimesh/exchange/gltf.py:343