Reduce unnecessary reliance and usage of numpy

[aobo-y]

This search gives us all the usage of Numpy in Captum https://github.com/pytorch/captum/search?p=1&q=numpy
Many of them are not necessary

• either empty imports
• or similar functionality can be replaced with pytorch

In order to reduce the unneeded dependency, we can investigate the current usages and remove/replace them when possible

[bilalsal]

Excellent find, @aobo-y!
Indeed, we noticed significant performance gain when we avoided numpy() while working on CCA and CKA.
You might need to pay attention to numerical precision issues, esp. to ensure that the unit tests continue to work meaningfully.

Thank you for bringing up this good proposal!

[aobo-y]

Thank you for the heads up @bilalsal !
I originally noticed this issue when helping @NarineK remove numpy for an internal use case (#714)
I will create a series of similar small PRs each of which contains only tightly-related changes, instead of one PR for all, so that we can review and discuss the removal/replacement case by case.

[NarineK]

@aobo-y can we close this issue since it has been addressed ?

[aobo-y]

@NarineK sure, I think we almost cleared the unnecessary cases.

However, there are still some tricky ones left. For example,

captum/captum/attr/_core/gradient_shap.py

Lines 359 to 363 in 4faf1ea

	rand_coefficient = torch.tensor(
	np.random.uniform(0.0, 1.0, inputs[0].shape[0]),
	device=inputs[0].device,
	dtype=inputs[0].dtype,
	)

is perfect scenario to directly apply torch.rand(...). But the change will require quite some extra efforts to fix many tests, because even with fixed seed during tests, torch's random gives different values from np. As you can imagine, it needs update of the expected values and relaxation in many "almost-equal" asserts.
I will create separate issues dedicated for them.