Handle models with divergent layer sizes

[dhiltgen]

The recent refactoring of the memory prediction assumed all layers are the same size, but for some models (like deepseek-coder-v2) this is not the case, so our predictions were significantly off.

Without the fix:

time=2024-06-18T11:03:42.708-07:00 level=INFO source=memory.go:303 msg="offload to metal" layers.requested=-1 layers.model=28 layers.offload=28 layers.split="" memory.available="[96.0 GiB]" memory.required.full="2.4 GiB" memory.required.partial="2.4 GiB" memory.required.kv="432.0 MiB" memory.required.allocations="[2.4 GiB]" memory.weights.total="1.6 GiB" memory.weights.repeating="1.4 GiB" memory.weights.nonrepeating="164.1 MiB" memory.graph.full="72.0 MiB" memory.graph.partial="72.0 MiB"

With the fix:

time=2024-06-18T11:02:47.707-07:00 level=INFO source=memory.go:309 msg="offload to metal" layers.requested=-1 layers.model=28 layers.offload=28 layers.split="" memory.available="[96.0 GiB]" memory.required.full="9.2 GiB" memory.required.partial="9.2 GiB" memory.required.kv="432.0 MiB" memory.required.allocations="[9.2 GiB]" memory.weights.total="8.4 GiB" memory.weights.repeating="8.3 GiB" memory.weights.nonrepeating="164.1 MiB" memory.graph.full="72.0 MiB" memory.graph.partial="72.0 MiB"

Partial fix for #5113 but we'll need additional graph updates...