sound: Implement AFMT_FLOAT support

Even though the OSS manual [1] advises against using AFMT_FLOAT, there
are applications that expect the sound driver to support it, and might
not work properly without it.

This patch adds AFMT_F32_LE|BE (as well as AFMT_FLOAT for OSS
compatibility) in sys/soundcard.h and implements AFMT_F32_LE|BE <->
AFMT_S32_LE|BE conversion functions. As a result, applications can
write/read floats to/from sound(4), but internally, because sound(4)
works with integers, we convert floating point samples to integer ones,
before doing any processing.

The reason for encoding/decoding IEEE754s manually, instead of using
fpu_kern(9), is that fpu_kern(9) is not supported by all architectures,
and also introduces significant overhead.

The IEEE754 encoding/decoding implementation has been written by Ariff
Abdullah [2].

[1] http://manuals.opensound.com/developer/AFMT_FLOAT.html
[2] https://people.freebsd.org/~ariff/utils/ieee754.c

PR: 157050, 184380, 264973, 280612, 281390
Sponsored by: The FreeBSD Foundation
MFC after: 1 week
Reviewed by: emaste
Differential Revision: https://reviews.freebsd.org/D47638