Add TieredAllocator for capacity-bounded on-chip/spill allocation

[fpedd]

The existing allocators all minimize a single unbounded pool. They cannot
answer the question a hardware team actually asks: given a fixed on-chip
memory, which buffers live there and which spill off-chip? TieredAllocator
fills that gap.

It models the two tiers within one address space: the fast memory occupies
offsets [0, capacity), every buffer that first-fit can place under that
ceiling stays on-chip, and the rest are packed contiguously above capacity
(the spill region). A buffer is on-chip iff its height (offset + size) is at
most capacity, so the result is an ordinary valid allocation and works with
validate_allocation, plot_allocation, and the benchmark harness unchanged.

TieredConfig exposes capacity (bytes; None means unbounded, the default) and
order, the priority buffers get for the fast tier (default largest size
first, keeping the bytes most expensive to spill on-chip). With capacity=None
the allocator reduces to first-fit in order, so it is constructible with no
arguments and usable directly by registry name "tiered_allocator".