[Breaking] Remove max_delta_step from Poisson.

[trivialfis]

User can use the parameter with same name max_delta_step in tree method to control leaf weights, no
need to manipulate hessian values.

[codecov-commenter]

Codecov Report

Merging #7267 (004fd0c) into master (d8a549e) will not change coverage.
The diff coverage is n/a.

@@           Coverage Diff           @@
##           master    #7267   +/-   ##
=======================================
  Coverage   82.64%   82.64%           
=======================================
  Files          13       13           
  Lines        4067     4067           
=======================================
  Hits         3361     3361           
  Misses        706      706           

---

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 d8a549e...004fd0c. Read the comment docs.

[RAMitchell]

It seems to me that this parameter is different from the other max_delta_step. It looks like it was put here to stop numerical instability problems. In particular the default value means the hessian cannot drop below 2 (assuming weights are 1.0). Can we be sure this change will not cause problems?

[trivialfis]

@RAMitchell It might cause problems in hessian as you mentioned due to small values. But having it might also cause trouble if the hessian is supposed to be small. It's a hyper-parameter that needed tuning.

To me, both max_depth_step serve the same purpose. The weight of the leaf is roughly this:

min(G/H, delta)

If you have a small hessian, G/H will be large, then max_delta_step in tree param chunks the output weight (G/H). On the other hand, the max_delta_step in the objective increases the hessian which is also to reduce G/H. So they are both trying to reduce the leaf weights.

We should use only one of them.