We only considered the audio already buffered through opensles engine,
so the calculation was incorrect.

Use samples rather than bytes for calculation, and a function rather than
hardcoding 2*2 all over the place when a size in bytes is needed.

Ubuntu bug #1038303

And simplify accordingly.

With hardware FPU, FI32 was only used within the mad decoder and
internally converted to FL32. With software FPU, FI32 did not work
properly as the S16N conversion was broken (at least since VLC1.1.0).

Apparently, mad generates sample values outside the [-1,+1] range, so
FL32 seems to be the most reasonable output format.

(On systems without hardware FPU, I would recommend disabling mad, and
using libavcodec for MPGA decoding.)

Previously, FPU detction was done before the variables for compiling (CC, CFLAGS, etc) were set. This broke on cross-compiles, because it was trying to use the default $(CC) to do the checks.

By doing the FPU detection after setting all the variables, FPU detection for cross-compiles works spectacularly.
Signed-off-by: Rafaël Carré <funman@videolan.org>

Close #6243

No functionnal changes */