[DRAFT] Allow manual and automatic memory/compute working together

[fedimser]

This PR allows to enable memory-compute architecture with LRU/LFU cache, and use manual Load/Store instructions. So that runtime automatically writes qubits to memory when not needed, but the user can force such a write to memory at any point.

TODO: this PR misses removing qubits from cache on release.

[msoeken]

I would remove all references to "hot" and "cold". Let's consistently use "compute" and "memory".

[msoeken]

I think that this is already too detailed and does not match with what I had in mind at our last meeting. I suggest a single intrinsic operation AssertComputeQubits(qubits : Qubit[]) (or something like that), with the following meaning:

• After the operation has been called, only qubits are in compute. Compute qubits that were present before the call and that are not in qubits are assumed to be moved back to memory. Qubits in qubits that were not compute qubits before need to be read from memory.
• Any gate that tries to access a qubit that is not explicitly in compute must fail, unless this is used in combination with dynamic memory/compute allocation
• There is no need to write back to memory, because this happens at the next call to AssertComputeQubits

This would require to understand how we want to use EnableMemoryComputeArchitecture. I suggest that if the compute capacity is a positive number, we assume dynamic memory compute (even though calls to AssertComputeQubits can be respected and may fail on insufficient capacity) and if we call it with 0 or a negative number (or any other kind of signature if that's better), there is no dynamic movement, but all compute qubits need to be pre-asserted with AssertComputeQubits.