Make summarize_tensor robust to non-float dtypes

[piercus]

Current implementation of summarize_tensor can raise secondary error while obfuscating the original error.

Actual

When doing some experiments, the stack was displaying a secondary error

Traceback (most recent call last):
  File "/home/pierre/dev/refiners/scripts/training/finetune-ldm-color-palette.py", line 15, in <module>
    trainer.train()
  File "/home/pierre/dev/refiners/src/refiners/training_utils/trainer.py", line 89, in inner_wrapper
    result = func(*args, **kwargs)
  File "/home/pierre/dev/refiners/src/refiners/training_utils/trainer.py", line 528, in train
    self.epoch()
  File "/home/pierre/dev/refiners/src/refiners/training_utils/trainer.py", line 512, in epoch
    self.step(batch=batch)
  File "/home/pierre/dev/refiners/src/refiners/training_utils/trainer.py", line 503, in step
    loss = self.compute_loss(batch=batch)
  File "/home/pierre/dev/refiners/src/refiners/training_utils/color_palette.py", line 234, in compute_loss
    prediction = self.unet(noisy_latents)
  File "/home/pierre/anaconda3/envs/refiners/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1518, in _wrapped_call_impl
    return self._call_impl(*args, **kwargs)
  File "/home/pierre/anaconda3/envs/refiners/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1527, in _call_impl
    return forward_call(*args, **kwargs)
  File "/home/pierre/dev/refiners/src/refiners/fluxion/layers/chain.py", line 279, in forward
    result = self._call_layer(layer, name, *intermediate_args)
  File "/home/pierre/dev/refiners/src/refiners/fluxion/layers/chain.py", line 273, in _call_layer
    raise ChainError(message) from None
refiners.fluxion.layers.chain.ChainError: RuntimeError:
   File "/home/pierre/dev/refiners/src/refiners/fluxion/utils.py", line 199, in summarize_tensor
    f"mean={tensor.mean():.2f}",

mean(): could not infer output dtype. Input dtype must be either a floating point or complex dtype. Got: Long

Expected

I'm expecting to get to orignal error in the stack

Which in my case was

Expected all tensors to be on the same device, but found at least two devices, cuda:1 and cuda:0! (when checking argument for argument mat1 in method wrapper_CUDA_addmm)

Solution

This PR implements a fix for this

[limiteinductive]

Seems good to me, important fix

[piercus]

Thanks for your feedback @limiteinductive.

I feel it worth unit testing summarize_tensor.
It will prevent any regression here (which might not be visible elsewhere in the CI).

[deltheil]

Thanks @piercus!