Some tuning

[lemire]

1. I have added filtering to the 32-bit and 64-bit test programs (exhaustivefloat32 and thoroughfloat64) so we can test just one algorithm.
2. I have tuned the string generation used for grisu_exact and teju_jagua. See below for results.
3. I changed how we can filter several algos... you do "-a myalgo1,myalgo2".
4. In the benchmarks, we now also benchmark the 'string only' conversion (see point 2).

Before (master):

grisu_exact                   :   19.19 i/B   360.70 i/f (+/- 0.0 %)      29.16 b/f     
teju_jagua                    :   16.56 i/B   311.26 i/f (+/- 0.0 %)      29.16 b/f   

This PR:

grisu_exact                   :   18.55 i/B   348.66 i/f (+/- 0.0 %)      28.36 b/f     
teju_jagua                    :   15.92 i/B   299.22 i/f (+/- 0.0 %)      28.36 b/f   

So slightly fewer instructions.

What my results seem to indicate is that we are heavily bounded on the string generation part. That is, printing out the characters is the slow part.

I think that teju_jagua fails exhaustivefloat32, but it did before too... so I do not think it is a new bug.

Interestingly, merely printing the characters from the mantissa can be slower than the whole dragonbox algorithm !!!

[lemire]

@jaja360 Merging.