Add bfp16 support

[L-roro]

This is a proposal to add bfp16 support to IRON. It is currently incomplete since I have not been able to make Peano work for the examples I have written, but they are apparently working with chess. I open this draft pr in order to get some feedback from the team.

The logic for the lowering is to simply transform the new block float type into an integer type of the appropriate length from the builtin mlir dialect. This lowering is done before the lowering of the object fifos into DMABDOps to avoid conflicts with functions calls that assume they are dealing with builtin types beyond that point in the pipeline.

There are currently no additional transformations being done and therefore management of the bfps is let at the discretion of the programmer. Currently, bfps are referenced as a unit from IRON (1 bfp represents 8 or 16 floats) to avoid dividing a bfp during data movement. This is in contrast to the AIE API which defines its vectors by referencing the number of floats represented. See bfp_conversion/kernel.cc for example and additional explanations.

Currently, different bfp types are instantiated uniquely in IRON (bfp16ebs8, bfp16ebs16). Nevertheless, they translate into a single MLIR type that encompasses all such types and identifies them through their AIE API name. I believe this is a reasonable compromise between syntax complexity and maintainability, that allows both for general optimizations for all block types or for individual ones if needed (despite not supporting any currently).