This PR improves string to double conversion by leveraging a C Wrapper for the C++ Eisel-Lemire ParseFloat algorithm.
This optimizes primarily our write workflows and we've measured an improvement of ~=8% on the achievable ops/sec.

Bellow you can find the top-down tree of cpu usage that made us focus on this optimization.

Top-Down tree of hotspots by CPU utilization for TS.MADD....

Please notice the 16.7% line, which we're optimizing on this PR.

Validating improvements with a tsbs dataset

on a quick look, using a single shard to test we move from 1.4M datapoints/sec ( 140K ops/sec of TS.MADD with 10 points ) to 1.53M datapoints.

According to Eisel-Lemire ParseFloat algorithm numbers, we should see a 4X improvement on the string to double, which means we would move from 17% CPU time on conversion to 0.25*17%=4.25% on best case scenario. I'm seeing an real improvement of ~8-9% on ingestion (meaning we're at 2X improvement ) :

###########################
Standalone on master branch
###########################
Summary:
loaded 120960000 metrics in 85.206sec with 1 workers (mean rate 1419615.30 metrics/sec)
loaded 12097000 rows in 85.206sec with 1 workers (mean rate 141973.27 rows/sec)

###########################
Standalone on fast.doubleparser branch
###########################
Summary:
2021/03/23 20:08:17 finished benchmark
loaded 120960000 metrics in 78.765sec with 1 workers (mean rate 1535699.71 metrics/sec)
loaded 12097000 rows in 78.765sec with 1 workers (mean rate 153582.67 rows/sec)