Add JSON model dump functionality documentation

[grace-lam]

I wrote a native XGBoost to Vespa converter that uses XGBoost's JSON model dump as input (vespa-engine/vespa#6577), so I added some missing documentation for the existing dump_format parameter in the CLI version.

[codecov-io]

Codecov Report

Merging #3600 into master will increase coverage by 0.1%.
The diff coverage is n/a.

@@             Coverage Diff             @@
##             master    #3600     +/-   ##
===========================================
+ Coverage     49.93%   50.04%   +0.1%     
  Complexity      188      188             
===========================================
  Files           172      172             
  Lines         13877    13877             
  Branches        457      457             
===========================================
+ Hits           6930     6945     +15     
+ Misses         6722     6707     -15     
  Partials        225      225

Impacted Files	Coverage Δ	Complexity Δ
src/data/sparse_page_raw_format.cc	41.5% <0%> (-22.65%)	0% <0%> (ø)
src/data/sparse_page_dmatrix.cc	92.3% <0%> (-0.6%)	0% <0%> (ø)
python-package/xgboost/compat.py	90.9% <0%> (+1.81%)	0% <0%> (ø)	⬇️
python-package/xgboost/core.py	83.41% <0%> (+4.61%)	0% <0%> (ø)	⬇️

---

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 725f4c3...f99f525. Read the comment docs.

[hcho3]

Thanks!