Add 16 and maybe 8 bit FP datatypes

[lstewart]

Problem Statement

CPUs and GPUs have added support for 16 bit and 8 bit floating point datatypes. SHMEM should support them.

Proposed Changes

Propose adding support for important new datatypes. BF16 and FP16 are candidates.
FP8 is a less likely possibility.

Support is easy for data movement APIs, Put, Get, Collect, Broadcast, Alltoall
Support for computational atomics is easy if there is hardware support.
Support for Reductions is a good question for discussion. I think it might be useful to have a new kind of reduction in which the target buffer is a different datatype than the source. One could do a sum-reduct from BF16 to FP32 for example, to retain more precision.

Impact on Implementations

Impact on Users

Adds new capabilities, no impact on existing applications.

References and Pull Requests

The Great 8 Bit Debate of Artificial Intelligence
BF16 vs FP16

[jeffhammond]

Table 5 of http://openshmem.org/site/sites/default/site_files/OpenSHMEM-1.6.pdf has 8 and 16 bit types so the only thing not supported already is reductions on the smaller floats.

Reductions of FP16, BF16 and FP8 should use FP32 accumulators. This requires a splat (copy+cast) but that's not going to be a bottleneck compared to the network reduction. Doing the first phase of the reduction sending the narrower type instead of FP32 is not a large savings, because all subsequent phases will send FP32.