Provide symmetric, color data preserving conversions between associated and
separate alpha for alpha values near and equal to 0.0 at single precision
floating point. Thus these symmetric conversions also augment associated alpha
to always maintain color information.

This is achieved by clamping the alpha used when computing the color
components when multiplying/dividing components between separate and
associated alpha to BABL_ALPHA_FLOOR and -BABL_ALPHA_FLOOR for values
near 0.0, the alpha value is copied unmodified between pixels; the main
improvement of this commit.

In the implementation BABL_ALPHA_FLOOR is 1/65536 = 0.00001526.. small
enough that it vanishing in rounding for 16bit integer alpha but large
enough to retain color data.

The following table illustrates what corresponding values are between the
two alpha encodings supported. We here usea BABL_ALPHA_FLOOR of 0.01 instead
of 0.00001526 to make it easier to see what happens.

```
"RGBA float"                     "RaGaBaA float"
separate alpha                   associated alpha
non-premultiplied                pre-multiplied alpha

    R       G      B      A         Ra     Ga     Ba     A
  10.000   1.000  0.100  0.000     0.100  0.010  0.001  0.000
  10.000   1.000  0.100  0.005     0.100  0.010  0.001  0.005
__10.000___1.000__0.100__0.010_____0.100__0.010__0.001__0.010__
  10.000   1.000  0.100  0.020     0.200  0.020  0.002  0.020
  10.000   1.000  0.100  0.200     2.000  0.200  0.020  0.200
  10.000   1.000  0.100  0.400     4.000  0.400  0.040  0.400

1000.000 100.000 10.000  0.000    10.000  1.000  0.100  0.000
1000.000 100.000 10.000  0.005    10.000  1.000  0.100  0.005
1000.000_100.000_10.000__0.010____10.000__1.000__0.100__0.010___
 500.000  50.000  5.000  0.020    10.000  1.000  0.100  0.020
  50.000   5.000  0.500  0.200    10.000  1.000  0.100  0.200
  25.000   2.500  0.250  0.400    10.000  1.000  0.100  0.400
```

GEGL lets each operation compute with it's preferred encoding - for some
operations like blurs this is associated alpha, for others like color
adjustments it is separate alpha, perhaps even in CIE Lab based encodings
rather than RGB. An earlier iteration of this approach was already in use in
babl making un-erase mode and similar features in GIMP-2.10 work correctly
after blurring, the previous implementation did not preserve the
alpha value.

Just the core of the implementation follows:

```
 #define BABL_ALPHA_FLOOR_F  (1.0f/65536.0f)

static inline float
babl_epsilon_for_zero_float (float value)
{
 if (value <= BABL_ALPHA_FLOOR_F)
 {
   /* for performance one could directly retun BABL_ALPHA_FLOOR_F here
      and dropping handling negative values consistently. */
   if (value >= 0.0f)
     return BABL_ALPHA_FLOOR_F;
   else if (value >= -BABL_ALPHA_FLOOR_F)
     return -BABL_ALPHA_FLOOR_F;
 }
 return value;  /* most common case, return input value */
}

static inline void
separate_to_associated_rgba (const float *separate_rgba,
                                   float *associated_rgba)
{
  float alpha = babl_epsilon_for_zero_float (separate_rgba[3]);

  associated_rgba[0] = separate_rgba[0] * alpha;
  associated_rgba[1] = separate_rgba[1] * alpha;
  associated_rgba[2] = separate_rgba[2] * alpha;
  associated_rgba[3] = separate_rgba[3]; /* direct copy */
}

static inline void
associated_to_separate_rgba (const float *associated_rgba,
                                   float *separate_rgba)
{
  float alpha = babl_epsilon_for_zero_float (associated_rgba[3]);
  float reciprocal_alpha = 1.0f / alpha; /* the conditional normally in
                                            this conversion to avoid
                                            division by zero is handled by
                                            epsilon_for_zero */
  separate_rgba[0] = associated_rgba[0] * reciprocal_alpha;
  separate_rgba[1] = associated_rgba[1] * reciprocal_alpha;
  separate_rgba[2] = associated_rgba[2] * reciprocal_alpha;

  separate_rgba[3] = associated_rgba[3]; /* direct copy */
}
```