Math: Replace exp() function with more accurate and fast version

[singalsu]

The function sofm_exp_int32() if replaced with new sofm_exp_approx()
function. It's not a direct replacement so the name is
changed. The input range changes from +/- 5 to +/- 8.
The wrapper function sofm_exp_fixed() with input range -16 .. +7.6
is updated to use the new function with a simpler range reduction
for large negative input values. All current exponent function
usage in SOF components is for the wrapper function that remains
compatible.
The function sofm_db2lin_fixed() is updated to handle larger
negative input decibels values made possible by the new more
accurate exponent function.
The accuracy in exponent function for input ranges from -8 to +8 or
-16 to 7.6 depending on used exponent function improves from
absolute error of 1 to 1000 ppm to constant less than 1 ppm.
The cycles count for exponent function calculate decreases from
about 369 max to 78 max on MTL platform for HiFi code version with
-O2 or -O3 optimization. The generic C version on MTL platform is
max 88 cycles.
The DRC component is a heavy user of the exponent function. The
saving in MTL build of DRC is 10.5 MCPS, it drops from 22.17 to
11.70 with profiler run:
"scripts/sof-testbench-helper.sh -x -m drc -p profile.txt"
The saving in MTL build of multiband DRC is 30.2 MCPS, it drops from
127.95 to 97.78 with profiler run:
"scripts/sof-testbench-helper.sh -x -m drc_multiband -p profile.txt"
This patch also updates the cmocka test. The test functions are
changed due to changed function name, changed achievable accuracy.
Also the random values test points vector is replaced fixed a
fixed linearly spaced vector for repeatable test and risk for
random failures.

[singalsu]

Still working with cmocka test cases, while the generic C and Xtensa HiFi versions of the exp functions are ready. Here's plots of accuracy (vs. C math library function) and cycles for existing HiFi version and new generic and HiFi version.

[singalsu]

Here's a cycles plot that includes cycles for C library float version. It's about 1800 cycles except for near zero input values.

[ShriramShastry]

Please address checkpatch space preference.
Over all the implementation looks sound to me.
Thank you

[singalsu]

Please address checkpatch space preference.

Ah that, it's an issue in checkpatch. It doesn't understand numbers in exponential format, e.g. 1.9073e-06.

[lgirdwood]

@singalsu good to merge ?

[singalsu]

@singalsu good to merge ?

Yep!