Add compatibility with skip_memory_metrics for mps device

[SunMarc]

What does this PR do

This PR makes mps device compatible with memory profiling integrated in Trainer (skip_memory_metrics=False). However, we can't get the information on the peak memory since torch.mps.max_memory_allocated() doesn't exist yet.
Fixes #27181
Tested on mac and I get the following output for a simple Trainer example:

{'train_runtime': 7.8136, 'train_samples_per_second': 24.572, 'train_steps_per_second': 3.072, 'train_loss': 5.870065689086914, 'init_mem_cpu_alloc_delta': 950272, 'init_mem_gpu_alloc_delta': 512, 'init_mem_cpu_peaked_delta': 16384, 'init_mem_gpu_peaked_delta': 'Not available', 'train_mem_cpu_alloc_delta': 38354944, 'train_mem_gpu_alloc_delta': 2304, 'train_mem_cpu_peaked_delta': 16384, 'train_mem_gpu_peaked_delta': 'Not available', 'before_init_mem_cpu': 351813632, 'before_init_mem_gpu': 0, 'epoch': 3.0}

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[ArthurZucker]

🤗 thanks you for this long due fix!