VarBuilder cleanup

[LaurentMazare]

This PR cleans up a couple rough edges of the VarBuilder, it tries to preserve the user-facing API so as not to be too much of an intrusive change.

• This makes the var builder extensible via the Backend trait. All the current implementations are in var_builder.rs but it's easy to move them/add new ones etc.
• There is a simplified version of the trait for the case when no extra argument is required, this should cover most use cases.
• The sharding api use the full trait rather than the simplified version, the user facing api has been made more in line with the "normal" version.
• get_or_init is renamed to get_with_hints so as not to be misleading on potential initialization (this decision is deferred to the backend).

Pending/drawbacks:

• More documentation to make the concepts easier to grasp.
• There are lots of explicit lifetime (and a phantom type to help with this). This is only required because of the safetensors backend and seems a bit like an abstraction leak, we should consider using some self-referential hack as it would clean up the var-builder code a lot.

Testing:

• Ran the multi-process llama example.
• Ran various examples (llama, falcon, yolo) as well as the tests.