Support AMP with TPUs

Nobody has claimed this yet.

### Description & Motivation

Lightning currently supports `accelerator="tpu", precision="bf16-mixed"`, but so far, this just sets the `XLA_USE_BF16` environment variable:
- Trainer: https://github.com/Lightning-AI/lightning/blob/master/src/lightning/pytorch/plugins/precision/xlabf16.py
- Fabric: https://github.com/Lightning-AI/lightning/blob/master/src/lightning/fabric/plugins/precision/xlabf16.py

> Side note: why does Fabric also move the data to bf16?

The XLA team added support for automatic mixed precision (AMP). `XLA:GPU` uses a `GradScaler` and the `autocast` context manager, whereas `XLA:TPU` just uses the latter: https://github.com/pytorch/xla/blob/c9f2d91a234cdaf91f0bbdb044ec94e297ac839a/test/test_train_mp_mnist_amp.py#L143-L147

### Pitch

Integrate `from torch_xla.amp import autocast, GradScaler`

The code would be very similar to the non-XLA AMP plugin: https://github.com/Lightning-AI/lightning/blob/master/src/lightning/fabric/plugins/precision/amp.py

This would likely replace our existing `XLABf16Precision` plugin with an `XLAMixedPrecision` plugin.

### Alternatives

**This was just merged upstream. It's likely very experimental**. I expect it will be released with PyTorch 2.1.

### Additional context

PR on PyTorch: https://github.com/pytorch/pytorch/pull/96370
PR on XLA: https://github.com/pytorch/xla/pull/5161

cc @borda @carmocca @justusschock @awaelchli @JackCaoG @steventk-g @Liyang90

Start by comparing the existing precision implementations in src/lightning/pytorch/plugins/precision/xlabf16.py and src/lightning/fabric/plugins/precision/xlabf16.py with src/lightning/fabric/plugins/precision/amp.py. Review the linked torch_xla AMP example and determine what integration is needed for TPU and XLA:GPU; done means the existing XLABf16Precision approach is replaced or extended with the requested AMP behavior.

Written by the indexing model from the issue text.