<atomic>: enable more atomic sizes, improve non-lock-free lock

Several things should be done with atomics with new ABI

• #843 enables atomic_ref<type_of_16_bytes_size>, but not atomic<type_of_16_bytes_size>, need to enable these as well
• Some believe that x64 CPUs without cmpxchg16b exist. Rule them out from vNext target hardware and use the instruction unconditionally.
• Make sure non-lock-free atomics use OS wait. This can be done by replacing inner lock with SRWLOCK or CRITICAL_SECTION, or by using of WaitOnAddress, etc. SwitchToThread or timed backoff, or pure active spinning are not good options
• Make sure atomic<T> with odd size, such as 3 bytes are padded up to next native atomic size. Make sure it is done only for atomic, not for atomic_ref.
• Consider eliminating pointer to SRWLOCK in atomic_ref and always locating it from pointer hash.

vNext note: Resolving this issue will require breaking binary compatibility. We won't be able to accept pull requests for this issue until the vNext branch is available. See #169 for more information.

Start by reading #843 and #169, then confirm that the vNext branch is available before beginning. Review the atomic and atomic_ref implementation and address each listed ABI, lock-free, padding, and pointer-hashing requirement; done means the vNext changes are covered without the rejected spinning approaches.