[Linter] Ability to disable some lints

🚀 The feature

There are several options to disable specific linters.

Option 1. Disable with linter-ignore: code

Pros:

• Similar to known syntax of various linters

Cons:

• Need to modify code of datasets to disable something

datapipe = datapipe.sharding_filter().shuffle()  # linter-ignore: shuffle-shard

Option 2. Global & Context disables

Pros:

• Can control datasets without modification of the code

Cons:

• Global might disable important errors
• Context requires additional indent
• Syntax feels weird
• Annoying to disable construct time linters (see below)

from torchdata import linter
linter.disable('shuffle-shard') # global
with linter.disable('shuffle-shard'): # context based
    dl = DataLoader2(...)

Option 3. DLv2 argument / ReadingService argument

Pros:

• Local to specific DataLoader
• Can control datasets without modication of the code

Cons:

• Syntax feels weird
• Some linters might trigger/not in various ReadingServices
• Annoying to disable construct time linters (see below)

dl = DataLoader2(dp_graph, [adapter], disable_lint = ['shuffle-shard'])

Option 4. DataPipe 'attribute'

Pros:

• Can be defined by DataSet developer or by the user
• Can impact construct time error handling

Cons:

• Syntax feels weird

datapipe = datapipe.sharding_filter().shuffle().disable_lint('shuffle-shard')

and/or (as we can have an adapter to do the same job)

dl = DataLoader(dp_graph,[DisableLint('shuffle-shard')], ...)

Personally, I prefer the last variant, but I'm open to discussion.

No files or tests are named. Start by tracing the linter entry points and how DataPipe, DataLoader2, and ReadingService options are handled; done means agreeing on one disable mechanism and implementing it with coverage for the relevant linting paths.