Register promotion: provide an option to limit the number of promoted elements

[ftynse]

Similarly to shared memory promotion, we may want to limit the number of elements promoted to registers. In particular, it should be less than the number of available registers of a given type. We may want to reuse the shared memory heuristic for picking tensor reference groups to be promoted.

Unlike shared memory, there is no point in reusing the same variables for different values since we don't control the variable<->register mapping anyway. Such reuses are likely to degrade the result of the downstream register allocator.