cast to float32 or float64 tensor when passing scalar to self.log

[MF-FOOM]

Bug description

#18686 made it such that all ResultMetric values are stored at float32 or higher precision.

However, values passed in as floats to self.log still get auto-casted to a low precision type (if that's the default precision) via __to_tensor. This means that even though the internal result metric representation is precise, the values can loose precision before even reaching ResultMetric.

I'd suggest we modify __to_tensor to use a function like this (as introduced in #18686) to determine and set the dtype:

def _get_default_dtype() -> torch.dtype:
    """The default dtype for new tensors, but no lower than float32."""
    dtype = torch.get_default_dtype()
    return dtype if dtype in (torch.float32, torch.float64) else torch.float32

What version are you seeing the problem on?

v2.1

How to reproduce the bug

No response

Error messages and logs

N/A

Environment

Current environment

#- Lightning Component (e.g. Trainer, LightningModule, LightningApp, LightningWork, LightningFlow):
#- PyTorch Lightning Version (e.g., 1.5.0):
#- Lightning App Version (e.g., 0.5.2):
#- PyTorch Version (e.g., 2.0):
#- Python version (e.g., 3.9):
#- OS (e.g., Linux):
#- CUDA/cuDNN version:
#- GPU models and configuration:
#- How you installed Lightning(`conda`, `pip`, source):
#- Running environment of LightningApp (e.g. local, cloud):

More info

No response

cc @carmocca

[MF-FOOM]

bump on this? i'm happy to implement, just want approval to go ahead with a pr

[carmocca]

Please go ahead @MF-FOOM! Sorry for the wait