Avoid recomputatation of `transformation` inside of `DisplacementField.warp`

Whenever one calls `warp` the transformation needs to be computed from the displacements.

https://github.com/BlueBrain/Atlas-Download-Tools/blob/60f0221916f4f9987a4728807be0b37934784cd0/src/atldld/base.py#L319

This operation requires `np.meshgrid` or `np.indices` and can be costly for big shapes. Anyway, the `DisplacementField` should be immutable and it should be enough to compute the transformation once.

### Some ideas
* LRU cache
* Explicit caching `e.g. (self._tx, self._ty)`
* Optionally pass it in the constructor (assuming it was computed already)

### Why relevant?
The primary example is when we download both the section and the expression image. We use the same `DisplacementField` to warp them, however, `transformation` is computed twice.

Start at src/atldld/base.py around line 319 and inspect DisplacementField.warp and how transformation is computed with np.meshgrid or np.indices. Compare repeated warps using the same field; done means an immutable DisplacementField reuses one computed transformation for both operations without changing warp results.

Written by the indexing model from the issue text.