perf: remove Mooncake warmup during preparation

[gdalle]

Mooncake preparation used to run the pullback or gradient once in order to compile the rule. This led to a x2 slowdown when used without preparation, because

DI.gradient(f, backend, x)

falls back on

prep = DI.prepare_gradient(f, backend, x)  # used to include one gradient call
DI.gradient(f, prep, x)

This PR removes the duplicate, so the first run of DI.gradient will be slower but only due to compilation.

Ping @penelopeysm who had asked about this

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 97.88%. Comparing base (ba8a2fa) to head (eadb50e).
Report is 1 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #772      +/-   ##
==========================================
- Coverage   97.88%   97.88%   -0.01%     
==========================================
  Files         128      128              
  Lines        7463     7460       -3     
==========================================
- Hits         7305     7302       -3     
  Misses        158      158              

Flag	Coverage Δ
DI	98.94% <100.00%> (-0.01%)	⬇️
DIT	95.20% <ø> (ø)

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[willtebbutt]

LGTM -- thanks for looking into this.