Cache: standardize cache interface

[gante]

What does this PR do?

In #27931, where the static cache was introduced, we noticed it had the following hard requirements:

1. The model instance holds the cache, as opposed to being a tensor passed around;
2. Each layer has its own cache, as opposed to a single cache for all layers.

This contrasts with previous implementations (e.g. DynamicCache). Given the hard requirements of the static cache, and the resulting benefits, this PR migrates the interface for all cache classes so as to match the static cache. As a result, the modeling code becomes slightly simpler 🤗

[gante]

Model correctness depends on #28937, rebasing with its contents

[gante]

Closed in place of #29180 (merge conflicts 🤷 )