fix(metal): GDN bfloat16, PA scheduler, error handling, MLX SDPA fixes

[emanueleDiVizio]

Summary

This PR fixes multiple correctness, performance, and stability issues encountered while running mistral.rs on Apple Silicon (M-series) with real multi-user inference workloads (Qwen3.5 MoE + Mixtral).

The changes focus on:

• Metal backend correctness (GDN + KV cache)
• Scheduler behaviour under load (PagedAttention)
• Robustness in concurrent serving scenarios
• MLX integration improvements for attention kernels

Several of these issues only surface under concurrent decode or long-running sessions.

Key changes

Scheduler (from upstream PRs #2031/#2034)

• Fix O(N²) thrashing in PagedAttention scheduler under mixed waiting/active workloads
• Introduce FCFS priority ordering to prevent starvation

GDN / Metal

• Fix dtype mismatch (bfloat vs bfloat16_t) in Metal kernels
• Add per-sequence fallback for concurrent decode when recurrent offsets diverge

Stability

• Replace panic on client disconnect with error handling
• Return error instead of panic on block allocation failure (race condition)

Performance / Features

• Increase Metal KV cache default max_seq_len (4K → 16K)
• Add optional MLX SDPA backend with Metal flash attention (head_dim=256 support)

Test plan

• Validated on Apple Silicon (M-series)
• Tested with Qwen3.5 MoE (GDN) and Mixtral
• Scheduler fixes verified under concurrent request load