BufferPool: make it user-controllable via eval() API

[shinaoka]

Summary

Currently BufferPool is always ON, created internally in EinsumNode::eval(), and not controllable by users. For small tensors, jemalloc/tcmalloc tcache is sufficient and the pool adds unnecessary complexity. For large tensors (>4MB), the pool avoids costly mmap/munmap.

Proposal

1. Make BufferPool public and accept Option<&mut BufferPool> in the eval API:

// With pool (user-managed, reusable across calls)
let mut pool = BufferPool::new();
node.eval_with_pool(&operands, &mut pool)?;

// Without pool (allocator handles everything)
node.eval(&operands)?;

2. This also enables sharing a single pool across multiple einsum calls.

Benefit

• Users can benchmark ON vs OFF to determine if the pool helps their workload
• Pool lifetime is user-controlled (can persist across calls or be dropped)
• Cleaner separation of concerns

Risk

Low. The pool is already passed as an argument internally (eval_node(..., pool, ...)). This just exposes it at the public API level.