Add variance calculation from FusedAdam optimizer states

[kwyss-nvidia]

Description

The state in the Adam optimizer provides moving averages for the first and second moments.
These can be combined to yield a gradient variance for a parameter.

This MR adds a method to the Adam optimizer to calculate a variance for a parameter.

Fixes # (issue)

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

Please list the changes introduced in this PR:

Add

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

[timmoon10]

Suggested change

	"""Return the unscaled state corresponding to the input `param` and `state_name`.
	Arguments:
	param (torch.nn.Parameter): One of parameters in this optimizer.
	state_name (string): Name of optimizer states, can be one of 'exp_avg', 'exp_avg_sq',
	and 'master_param`.
	"""
	"""Estimate the gradient variance based on moment estimates."""

[timmoon10]

We could make this test more robust by manually computing exp_avg and exp_avg_sq outside of the optimizer. There's some complicated dtype-specific logic within get_unscaled_state, so I don't really trust it.