remove old (obsolete?) "compressed" floatTransfer

• I don't think that it is being used anywhere, anymore. Could suspect that we wouldn't necessarily want rounding differences between two sides of a processor field

Opinions @andy, @Mattijs ?

changed milestone to %v2506

added code style label

assigned to @mark

• plus: if your case is communication dominated allows a bit speedup
• minus : for AMG the lowest level is almost purely processor comms
• minus : for most of its use (halo swapping of residuals) the numbers are very small so will probably not converge in single precision
• question : remember half precision preconditioning? I.e. precision in some cases does not matter.
• minus : is only supported for halo swaps, not many-to-many (e.g. PstreamBuffers etc.).
• minus : precision truncation might give problems in e.g. geometry processing (we assume opposite processor face orientation)

Me : happy for it to disappear.