Draft "Unsafe fp optimizations" project description.

[Joern Rennecke]

> My real point here was not so much to get an accurate statement of how
> often this could cause trouble, but rather to avoid a statement that I
> thought was far too pessimistic and scary :-)

I think we should describe it in terms of safe value ranges.  E.g. if
you use only half the exponent range (which is like having one less
exponent bits), you are safe from overflows for transformations like
a/b/c -> a/(b*c)

Of course, higher order transformations eat more exponent bits.