Fix amplitude_to_DB numerical instability

[mthrok]

Recently we discovered that torch.log10 behaves differently on different CPUs when the input tensor is float32. pytorch/pytorch#37833 Note: This is not a random behavior.

Due to this, we have some tests that passes on some machines but fail on other machines.
A proposed fix for this is to use float64 type before passing the tensor to torch.log10 here, and convert back to float32 if the input type was float32

Once this is completed, we could re-enable the following tests;

1. librosa_compatibility_test
2. test_batch_consistency

For 2. See the commit b499840 and revert the atol to 1e-5.

You can run these test with the following commands;

pytest test -v -k batch_mfcc
pytest test -v -k basics3

In addition to that, all the existing test that uses amplitude_to_DB, should pass.

pytest test -v -k amplitude_to_DB

[Micheladar]

I'll work on it.

[kshitij12345]

@Micheladar Hi, are you working on it?

[discort]

@mthrok
Is still relevant issue? If so, could you please update the description. It seems there are no mentioned tests

[mthrok]

Resolved via #1436 #1439

@kshitij12345 @discort sorry for not responding earlier. The test has been refactored and refined. This issue is no longer relevant.