Add Apple MPS NF4 benchmark

[hamedrabah]

Summary

Adds a standalone Apple MPS benchmark for the now-supported NF4 path. It
measures:

• NF4 quantization and dequantization median latency
• packed tensor plus absmax metadata size relative to fp32
• round-trip quantization error
• Linear4bit latency and output error relative to dense fp16

The benchmarking README includes the command and the script exposes tensor
size, layer size, warm-up, and repetition controls. Results are reported rather
than asserted so the benchmark remains useful across Apple Silicon variants.

Validation

Tested on Apple M4 / macOS 15.7.9 with PyTorch 2.13.0 against the current
checkout (bitsandbytes 0.50.2.dev0):

$ PYTHONPATH=. python benchmarking/mps/nf4_benchmark.py --repeats 3 --warmups 1
device: Apple MPS (arm64)
torch: 2.13.0; bitsandbytes: 0.50.2.dev0
NF4 packed+absmax storage: 0.141x of fp32 (16.0 MiB -> 2.2 MiB)
quantize median: 7.950 ms
dequantize median: 1.110 ms
round-trip MAE: 0.072869; max error: 0.654016
dense fp16 GEMV median: 0.870 ms
NF4 Linear4bit median: 1.996 ms
dense-vs-NF4 output MAE: 3.361328

$ ruff check benchmarking/mps/nf4_benchmark.py
All checks passed!

$ ruff format --check benchmarking/mps/nf4_benchmark.py
1 file already formatted

The existing 4-bit test selection was also exercised on MPS: 252 passed, 36
skipped, 60 xfailed, and 426 deselected.