[v3] Elevate codec pipeline

[d-v-b]

This PR moves the CodecPipeline data structure off the ArrayMetadata classes and instead localizes it higher in the stack.

design for array metadata classes

In zarr-python, we have ArrayMetadata classes that are expressly designed to model the contents of zarr metadata, e.g. zarr.json or .zarray (and these classes should not do anything else). Designing the ArrayMetadata classes around this goal is very deliberate, largely due to lessons we learned in the v2 codebase, where the zarr.Array API was mixed together with the a model of the .zarray JSON document, and this led to a variety of problems.

we are breaking that design in v3

With that in mind, note that the v3 spec defines that the codecs attribute of zarr.json is a JSON array of Codec objects. But in the current version of the v3 branch, ArrayV3Metadata.codecs has the type CodecPipeline, but CodecPipeline is not a collection of Codec objects; instead, it's an object with a lot of attributes and methods. Putting such an object under the ArrayV3Metadata.codecs key violates the design principle of the ArrayMetadata class.

one way to fix it

But the fix is relatively simple: we remove CodecPipeline from ArrayV3Metadata and instead push the responsibility for creating instances of that class to objects that consume ArrayMetadata. That's what this PR does.

Specifically, I make CodecPipeline an attribute of AsyncArray; that attribute is created on initialization, by calling a function on the metadata.codecs attribute. Sharding codecs also use CodecPipeline; in this PR, they create the CodecPipeline when needed instead of keeping it as an attribute. If we don't like this, we could consider adding a _codec_pipeline attribute to the sharding codecs, as long as it's clear that this attribute is not part of the JSON serialized form of the codec.

I made a few other changes in the process, including changing the signature of the Codec.validate method to explicitly take the parameters that it needs (shape, dtype, chunk_shape), instead of an ArraySpec object that includes extra fields (specifically, ArraySpec.attributes) that no codec will ever need for validation. If that's unpopular in this PR we can spin it out into its own PR or just not do it, but I think this change is an improvement.

@normanrz this touches a lot of code you authored so I will be very interested in getting your feedback here

[d-v-b]

the changes here would address some of the problems described in #1913

[normanrz]

I am not convinced that the class structure needs to match the json structure one-to-one. That is why we have the overridable from_dict and to_dict classes. We should be pragmatic, in particular, in places where we can make the user API more convenient.

I am fine with the changes to the CodecPipeline and can live with the changed validate signature. I am not happy with forcing the users to supply typesize and shuffle for the BloscCodec.

[d-v-b]

@normanrz i reverted the changes to the blosc codec

edit: feel free to ignore this PR until I fix the merge conflicts

[d-v-b]

the pre-commit action is flagging a mypy error that a) I can't replicate locally and b) is in a file that this PR did not touch, so I'm going to merge this despite the pre-commit failing.