Be more careful to not create ticks earlier than needed, as that's quite
costly.

On the following benchmark

from matplotlib import pyplot as plt
from time import perf_counter
fig = plt.figure()
for i in range(3):
    print(i)
    start = perf_counter()
    fig.subplots(10, 10)
    print(perf_counter() - start)  # creation time
    start = perf_counter()
    fig.clf()
    print(perf_counter() - start)  # clearance time

this speeds up axes creation from ~1.4s to ~0.5s and figure clearance(!)
from ~2.6s to ~0.3s.

Note that a similar idea was already implemented in c935965, which sped
up creation from ~2.6s to 0.8s and clearance from ~2.2s to ~1.8s, but
obviously the whole thing is brittle and got mostly lost at some
point...

Would be nice to track performance regressions (semi)automatically.

attn @timhoffm as you initially had this idea in #9727

(#16066 made this a bit more of a pain than expected)

edit: hmmm, looks like this just transfers the cost to draw time, will investigate more...

PR Summary

PR Checklist

• Has Pytest style unit tests
• Code is Flake 8 compliant
• New features are documented, with examples if plot related
• Documentation is sphinx and numpydoc compliant
• Added an entry to doc/users/next_whats_new/ if major new feature (follow instructions in README.rst there)
• Documented in doc/api/api_changes.rst if API changed in a backward-incompatible way