JSON parse gives wrong Float64 result

[hanabi1224]

Repro steps
Below code show difference between from_json and to_f64. (result of to_f64 is consistent with other languages while from_json is not)

num = "43.474709000000132"
puts Float64.from_json(num)
# 43.474709000000125
puts num.to_f64
# 43.47470900000013

Crystal compiler version

Crystal 1.0.0 [dd40a24] (2021-03-22)
LLVM: 10.0.0
Default target: x86_64-unknown-linux-gnu

[HertzDevil]

Should be fixed by #10447

The root issue is that Float64.from_json tries to obtain the floating-point value from 43474709000000132 / 1e15, which in turn calls 43474709000000132.to_f64 / 1e15.to_f64. The significand is inexact and rounds off to 43474709000000128.0 (it is halfway between this value and 43474709000000136.0, and the ties-to-even rule chooses this value), so the division becomes inexact. If you do the same division in other programming languages you could get the same incorrect results.

[straight-shoota]

Related to #10419