Implement fast-math Version of ReverseDouble with +ffi

[julmb]

3. I wonder if it is worth adding a "fast-math" version of ReverseDouble that achieves better performance at the expense of not being correct with respect to special IEEE floating-point values. This would be more work, but it would give a better migration story for anyone whose applications' performance might be negatively impacted by these changes.

Originally posted by @RyanGlScott in #105 (comment)

[julmb]

This is the relevant position in the code for this issue:

ad/cbits/tape.c

Lines 90 to 108 in 18efa2c

	double v = buffer[idx + pTape->offset];
	// TODO: if we do not care about handling IEEE floating point special values (NaN, Inf) correctly
	// then we can skip the rest of the loop body in case v == 0
	// see also https://github.com/ekmett/ad/issues/106
	int i = pTape->lnk[idx*2];
	if (i >= 0)
	{
	double x = v * pTape->val[idx*2];
	if (x != 0) buffer[i] += x;
	}
	int j = pTape->lnk[idx*2 + 1];
	if (j >= 0)
	{
	double y = v * pTape->val[idx*2 + 1];
	if (y != 0) buffer[j] += y;
	}

[cartazio]

Did including this check create a measurable change in any large examples? I forget if we discussed this before

[RyanGlScott]

@julmb has some large ad-based projects that incur a slight performance penalty because of this, if I understand #105 (comment) correctly.

[julmb]

Did including this check create a measurable change in any large examples? I forget if we discussed this before

Yes, in my application this would result in a performance improvement of around 20%.