[feature] CPU/Metal (device-agnostic) PagedAttention to complement the CUDA kernels from #3655

### Context

#3655 added paged flash-attention kernels in `candle-flash-attn`, resolving the
CUDA side of #3649. Those kernels are CUDA-only (the crate is built behind the
`cuda` feature), so there is currently no paged-attention path on CPU or Metal.

### Problem

Code that wants to use block-paged KV caches today must branch between the CUDA
flash-attn path and "something else" on CPU/Metal, where no paged primitive
exists. This hurts portability for tests, local development on Apple hardware,
and CPU fallbacks.

### Proposal

Add a device-agnostic paged-attention primitive (CPU + Metal, also runnable on
CUDA) as a complement to — not a replacement for — the fused CUDA kernels from
#3655:

- `paged_attention(q, k_cache, v_cache, block_tables, context_lens, block_size,
scale, alibi_slopes) -> Tensor`, implemented with plain candle tensor ops
(gather + dense attention), so it runs on every backend.
- `reshape_and_cache(...)` to write new K/V into a block-paged cache by slot
mapping.
- A small block-pool allocator for the KV-cache block infrastructure.

This would live in `candle-nn` (alongside the existing reference attention
variants) and serve as a portable fallback and a correctness reference for the
CUDA kernels.

### Notes

A working pure-candle reference implementation already exists (MHA/GQA/MQA +
optional ALiBi, validated on CPU against dense attention) and can be contributed
as a starting point. Happy to open a PR if this direction is welcome.

Relates to #3649, complements #3655.

Start in candle-nn alongside the existing reference attention variants and review the described pure-Candle implementation for MHA, GQA, MQA, and optional ALiBi. Implement the device-agnostic paged_attention, reshape_and_cache, and KV-cache block-pool pieces, then validate the portable path against dense attention on CPU and ensure it can run on Metal and CUDA.

Written by the indexing model from the issue text.