Support different cache types for each layer. Gradient computations when all or some caches are of type `DenseKVCache`

**Is your feature request related to a problem? Please describe.**

Currently, `DenseKVCache` cannot be used. This is because we need a fixed cache length and assume the buffers will be of this size most of the time. Also, it is unclear whether different cache types per layer are supported.

**Describe the solution you'd like**

I can use `DenseKVCache` for some or all layers. At any given point, memory only scales with the current cache size. Updates of the cache are of the `cat` type only.

This may be simple. We also want to make sure that different cache types per layer are supported.

No files, tests, or entry points are named. Start by locating the cache implementation and the gradient-computation path, then determine how cache types are selected per layer. Done means DenseKVCache works for some or all layers, memory follows the current cache size, and cache updates use concatenation.

Written by the indexing model from the issue text.