[PyTorch] Delegate delayed-scaling updates to RecipeState

[pggPL]

Description

Extract the non-behavioral recipe-state refactor from #34. FP8GlobalStateManager continues to own update timing and iteration over the existing delayed-scaling global buffers, while the corresponding RecipeState subclass owns the update algorithm.

This does not change the global buffer registry, update cadence, or make non-delayed recipes stateful.

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

• Extract recipe-to-RecipeState resolution into RecipeState.class_for_recipe.
• Reuse that mapping for state construction and global update dispatch.
• Move the existing delayed-scaling concat/reduce/scale-update implementation into DelayedScalingRecipeState.reduce_and_update_global_state.

Test plan

• pytest -q tests/pytorch/test_recipe.py plus targeted delayed-scaling CustomRecipe and fuser tests: 58 passed, 92 skipped.
• python -m pylint transformer_engine/pytorch/quantization.py: 10.00/10.
• Forced unfused module tests reproduce the same 12 failures on an unmodified main; the fuser variants pass in both cases.

Checklist:

• I have read and followed the contributing guidelines
• The functionality is complete
• I have commented my code, particularly in hard-to-understand areas
• No documentation changes are required for this internal refactor
• My changes generate no new warnings
• Existing tests cover the refactored behavior
• New and existing relevant unit tests pass locally with my changes