serve: auto-select metal-flash-sdpa attention on MPS
When `transformers serve` runs on Apple Silicon (`--device auto` or `mps`)
with `kernels` installed and no explicit `--attn-implementation` flag, default
the attention to `kernels-community/metal-flash-sdpa` instead of plain SDPA.
On the 100-sample gsm8k benchmark (Qwen2.5-0.5B-Instruct, MPS fp16,
generate_batch) it's a 1.66x throughput improvement (158 -> 256 tok/s) with
token-for-token parity for greedy decoding. Users who don't want it can opt
out with `--attn-implementation sdpa`.
Help text on the `--attn-implementation` flag also now lists the kernels-hub
syntax explicitly.
