Add two new operations to StgPrimFloat.c:

	__int_encodeFloat
	__int_encodeDouble

for encoding floats/doubles from small integers.  This avoids having
to convert small integers to large ones before an encodeFloat
operation, and fixes the two cases of slowdown in nofib after the
small integer changes.

Also:
	- remove encodeFloat and decodeFloat as primops
	- use foreign import for encode{Float,Double} and
	  the various isNaN etc. ccalls in PrelNumExtra.