Experimental reservation aware resource adapter and reservation

[nirandaperera]

[C++] Add experimental reservation-aware memory resource adaptor

Related to rapidsai/rapidsmpf#1142

Callers could only discover that a memory budget was exhausted at allocation time, with no way to claim a slice of it up front. This adds cucascade::memory::experimental, where you reserve bytes against a limit and then allocate through that reservation, so an over-budget request is refused at reserve() rather than failing deep inside a cudf call. Ported from rapidsmpf and extended so that a single reservation type covers device, host, and host+device memory.

• reserve(size, allow_overbooking) yields a memory_reservation that charges each allocation against its balance() and refunds the unspent remainder once the last reference dies. Overbooking is opt-in: permitted, the full size is granted and the shortfall reported by overbooking(); refused, a zero-sized reservation comes back and throws rmm::out_of_memory on first use.
• Three adaptors — device_adaptor, host_adaptor, and host_device_adaptor — forward their upstream's accessibility to every reservation they grant, so a device adaptor cannot hand out host-usable memory.
• memory_reservation is one type no matter where its memory lives. To hand it to cudf or RMM, project it with as_device(), as_host(), or as_host_device(); each throws cucascade::logic_error when the reservation lacks that accessibility, and accessibility() lets callers branch without exceptions.
• limit(), current_allocated(), total_reserved(), and available() expose the budget, with get_main_record() for peak and allocation-count statistics. The adaptor is itself a memory resource, so allocations made directly through it are still tracked against available(), just not capped.

device_adaptor adaptor{any_device_resource{rmm::mr::cuda_memory_resource{}}, 1GB};

auto res = adaptor.reserve(4MB, allow_overbooking::NO);

rmm::device_buffer buf{1MB, stream, res.as_device()};
assert(res.balance() == 4MB - 1MB);

auto table = cudf::groupby(..., stream, res.as_device()); // all allocations are capped against res

// The adaptor is a resource too: tracked against available(), but not capped.
rmm::device_buffer scratch{1ull << 20, stream, adaptor};