rand BFloat16 sampling without conversion

[milankl]

fixes #71

Technically randn, randexp should be written directly for BFloat16 too to avoid the rounding errors from conversion but they are 1) more complicted (see ziggurat algorithm for randn) 2) randn is continuous so rounding from conversion might be fine, randexp doesn't seem to produce very small numbers (<<1e-10) let alone 0 in practice for Float32/64, so rounding because of the conversion might be fine here too

Edit: I was wrong on randexp being complicated, from wikipedia

[ViralBShah]

@oscardssmith should we merge?

[oscardssmith]

This implementation is slightly less precise than necessary, why not use the same one used by Float16? (e.g. BFloat16(Float32(rand(r, UInt8)) * Float32(0x1.0p-8))

[JeffreySarnoff]

just saw this -- thank you