Early-convert BenchmarkGroup names to JSON-safe

[vtjnash]

It turns out we're missing a large chunk of tuning, since the JSON file forces all keys to be string, but we expect keys to be semi-rich types (for rich filtering). Compromise instead and coerce some types to string immediately and parse-eval other strings back to basic types (string, number, and tuples thereof).

[codecov-io]

Codecov Report

Merging #193 (c993884) into master (2e2e6ef) will decrease coverage by 2.05%.
The diff coverage is 70.49%.

@@            Coverage Diff             @@
##           master     #193      +/-   ##
==========================================
- Coverage   87.77%   85.71%   -2.06%     
==========================================
  Files           6        6              
  Lines         687      714      +27     
==========================================
+ Hits          603      612       +9     
- Misses         84      102      +18     

Impacted Files	Coverage Δ
src/serialization.jl	84.81% <64.70%> (-15.19%)	⬇️
src/groups.jl	93.37% <77.77%> (-3.15%)	⬇️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 2e2e6ef...c993884. Read the comment docs.

[vtjnash]

The de-JSON recover-y should also incidentally fix the awkward doubled-quoting of the reports in Nanosoldier such as https://github.com/JuliaCI/NanosoldierReports/blob/master/benchmark/by_date/2021-02/17/report.md

[ararslan]

Makes sense I think, though it's been ages since I've thought about the internals here.