Add configurable normalization epsilon for NTXentLoss

When training with half-precision I noticed that normalization in NTXentLoss can give NaN values.

in forward method, there is a code:

        # normalize the output to length 1
        out0 = nn.functional.normalize(out0, dim=1)
        out1 = nn.functional.normalize(out1, dim=1)

It uses torch.nn.functional.normalize function with default 1e-12 epsilon, what gives 0 for half precision. As a result we have division by zero and NaN in output.

The way to solve it is to add optional normalization epsilon parameter in NTXentLoss initializer and use it when calling torch.nn.functional.normalize function.

Please let me know if there is any mistake in my understanding.
If it's okay for you, I can propose a pull request.

Start by locating NTXentLoss and its forward method, then inspect the existing torch.nn.functional.normalize calls and related tests. Reproduce the half-precision NaN behavior if possible, and verify that a configurable epsilon is used for both normalizations while preserving the current default behavior.