add hexfloat (%a) support for printf

[simonbyrne]

This should match the behaviour of c99 standard (page 291).

[StefanKarpinski]

This is quite impressive. I should probably give it a serious read-through, but it looks overall good.

[simonbyrne]

It is based on the %e implementation, though it needed a few tweaks to deal with the auto-truncation behaviour.

The spec doesn't say how to choose the leading digit. I've implemented the gcc behaviour which uses a leading 1 most of the time (though we differ on handling of subnormals):

julia> @printf "%a" 9.0
0x1.2p+3

MPFR on the other hand appears to choose the leading digit to match the integer representation:

julia> @printf "%a" big(9.0)
0x9p+0

Not sure if it matters?

[StefanKarpinski]

Ouch, that's annoying that they don't agree. Printing the same value in different types should really produce the same result :-. Not sure how best to address that though.

[simonbyrne]

From what I can tell, gcc tries to closely represent the underlying bit structure: hence the leading 1 for most numbers, and for subnormals it prints things like 0x0.0000000000001p-1022.