This pull-request simplifies the tq.Transfer type, and removes the usage of the mutable Transferable interface. Additionally, it enables us to remove *api.ObjectResource from the tq's exported API.

Previously, we used three main types:

1. Transferable - an interface{} to cover downloadable and uploadable LFS pointers.
2. *api.ObjectResource - A type to hold values from the API.
3. *transfer.Transfer - A combination of Transferable() and *api.ObjectResource. Used in adapter implementations to grab values both from the API, and values pertaining to the local repository.

The bummer is that the Transferable is cached in a map[string]Transferable on the *TransferQueue for the entire transfer, which can incur some pretty high memory consumption. unsafe.Sizeof suggests that the we're storing much more than 56 bytes per Transfer, not including the size of the inner-structs. My calculations estimate about ~128bytes per object (for two actions (32 bytes/action), 4 bytes for the pointer), and no error).

To simplify, let's store the API data when we need it, and ignore it when we don't. For a local cache, all we need is the Name, Path, Oid and Size, which compacts to 56 bytes as a maximum (we can remove 16 bytes/transfer by dropping the Name field). This is, on average, 68.7% less memory consumption per unique Transfer.

In addition, the code gets simpler, since we only need to care about the objects retuned back from the API when we're sending them to the adapters.

On top of all of that, we got rid of the mutability in SetObject(), narrowed down our internal use of the *api.ObjectResource type, and only have one remaining external use of the api package in tq.

Since this PR is on the larger side, here are the important groups of commits:

1. e18044f: remove Transferable interface
2. 7f26137: simplify (q *TransferQueue) Add() signature.
3. 94b2d0d...5fc2344: various 💅 and 🍬 to clean things up

/cc @git-lfs/core