[JAX] Decouple Recipe and ScalingMode

[jberchtold-nvidia]

Description

Currently the recipe and scaling mode are coupled in the TE/JAX extension. This is okay for the current recipes, such as delayed scaling, current scaling, and MXFP8 block scaling, as there is only a single scaling mode used per recipe. However, for the DeepSeek recipe this assumption no longer holds. For the DeepSeek recipe we will need 1x128 1D block scaling for inputs and 128x128 2D block scaling for weights. As a result, we need to decouple the two concepts of recipe and scaling mode.

This PR only decouples the recipe and scaling mode, it does not implement the DeepSeek recipe.

Type of change

• Documentation change (change only to the documentation, either a fix or a new content)
• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• Infra/Build change
• Code refactoring

Changes

• Define UsageContext which defines the context in which a quantizer will be used (e.g. is the quantizer used for x, kernel, or grad)
• Add RecipeManager classes that provides recipe-specific functionality for quantization
• Replace QuantizeConfig.SCALING_MODE with QuantizeConfig.RECIPE_MANAGER and update QuantizeFactory to use the recipe manager instead

Checklist:

• I have read and followed the contributing guidelines
• The functionality is complete
• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• I have added tests that prove my fix is effective or that my feature works
• New and existing unit tests pass locally with my changes

[jberchtold-nvidia]

/te-ci L0

[jberchtold-nvidia]

/te-ci L0

[phu0ngng]

Hi, why do we need a new class just to wrap around an enum?

[phu0ngng]

I have a look at this QuantizerParams and on how it is used, and I would prefer not to have this class for the following reasons:

1. Whenever we need to query the scaling_mode or q_dtype or q_layout info, instead of

QuantizeConfig.RECIPE_MANAGER.get_quantizer_params(UsageContext(UsageType.X)).scaling_mode

We could have simply done

QuantizeConfig.RECIPE_MANAGER.get_scaling_mode(UsageType.X)

It is way simpler for other people to follow and make a contribution later.
2. With this QuantizerParam, we add one more level of object inside the Quantizer, which does not give any benefits. For the Quantizer create, we could do

q_x = QuantizerFactory.create(RECIPE_MANAGER.get_scaling_mode(UsageType.X),
                                                    RECIPE_MANAGER.get_quantize_dtype(UsageType.X),
                                                    RECIPE_MANAGER.get_quantize_layout(UsageType.X),
                                                    **args_x)

[phu0ngng]

Why don't we merge QuantizeConfig and RecipeManager into a single class?
I don't see a clear need for them to exist separately.