ideas for improving the store API

[d-v-b]

I've encountered a few pain points in our store API, which I will demonstrate with the following code snippet:

import zarr
from zarr.core.sync import sync
from zarr.storage import MemoryStore
import json

store = MemoryStore({})
zg = zarr.create_group(store)

# read zarr.json as JSON
# this process sucks
zmeta = json.loads(sync(store.get('zarr.json')).to_bytes())

IMO something like this would be much better:

import zarr
# stores are top-level exports
store = zarr.MemoryStore({})
zg = zarr.create_group(store)
# equivalent
zmeta = store.get_json('zarr.json')
zmeta = (store / 'zarr.json').get_json()

Lack of synchronous methods for IO

Problem:
Store IO is async, with no sync methods. Users must use private API (zarr.core.sync.sync) to invoke store.get in a synchronous context like a debugger or interpreter.

Solution:
Add synchronous methods for all the async methods. Ideally we would have get and get_async but this would be breaking as get is async today. A non-breaking change could be to add get_sync. But I think the proposed get_bytes and get_json methods described below can also solve this problem.

Buffer abstraction is a stumbling block

Problem:
Stores return buffer objects, which means you have to call to_bytes() to get the bytes.

Solution:
Add a get_bytes(*, path: str = '') method to the store that synchronously retrieves the object at path, and decodes the result as bytes. I think we would also get a lot of value from a get_json method.

StorePath / Store distinction is needless

Problem:
We have both Store and StorePath classes. Two classes instead of one adds needless complexity to the codebase.

Solution:
Fold all the StorePath logic into the Store class definition, and deprecate StorePath.

The StorePath definition is very simple -- it's just a store instance, and a string, and some methods that combine these 2 things. There is no reason for two separate classes here.