hi-mlx expert_pool: close fd leaks, fix mmap read overflow

[davidrhodus]

Summary

Two soundness/resource fixes in crates/hi-mlx/src/expert_pool.rs (MmapShard):

1. File-descriptor leak. MmapShard::open leaked both fds via std::mem::forget — the File backing the mmap and the F_NOCACHE direct fd — and Drop never closed the mmap fd. Each opened shard permanently consumed two descriptors, a slow leak across repeated model loads. Both fds are now owned via into_raw_fd and closed exactly once in Drop after munmap (no leak, no double-close), and the confused stale comment is removed.

2. Out-of-bounds read via integer overflow. MmapShard::read computed end = start + len without a checked add, so a huge offset/length pair could wrap past the end > self.len bounds check and read out of bounds. Now uses checked_add and rejects overflow.

Also prefixes the now-unused pad_attention_bias offset param in models.rs (the cache bound is correctly derived from kv_len - l per the in-loop rationale) to silence the warning.

Verification

• cargo test -p hi-mlx --features mlx --lib expert_pool: 9/9 pass.
• The full hi-mlx GPU suite is environmentally blocked in this sandbox (Metal toolchain can't build mlx.metallib) and fails identically on the base — no regressions from these changes.
• cargo check --workspace: clean.

Note: these changes are already on main (committed as 1513d46 during the review sweep); this branch carries the same fix isolated for review.

[davidrhodus]

These changes are already on main (committed as 1513d46 during the review sweep), so this isolated review branch is redundant. Closing without merging.