Fix performance of flatten! for large summations

[odow]

When fixing #4196 I didn't do a good job and introduced a pretty bad performance
regression for problems with large summations.

Script

using JuMP
function main(N)
    model = Model()
    @variable(model, x[i in 1:N])
    @expression(model, sum(100 * (x[i+1] - x[i]^2)^2 + (1 - x[i])^2 for i in 1:N-1))
end

v1.30.0

julia> @time main(10^5);
  0.430789 seconds (10.40 M allocations: 483.358 MiB, 67.91% gc time, 0.01% compilation time)

julia> @time main(10^5);
  0.426143 seconds (10.40 M allocations: 484.103 MiB, 63.67% gc time)

master

julia> @time main(10^5);
 12.954410 seconds (9.60 M allocations: 455.889 MiB)

julia> @time main(10^5);
 12.973724 seconds (9.60 M allocations: 457.638 MiB)

This PR

julia> @time main(10^5);
  0.208661 seconds (9.40 M allocations: 455.111 MiB)

julia> @time main(10^5);
  0.193844 seconds (9.40 M allocations: 458.643 MiB)

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 99.96%. Comparing base (55744bb) to head (d98f847).

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #4264      +/-   ##
==========================================
+ Coverage   99.95%   99.96%   +0.01%     
==========================================
  Files          43       43              
  Lines        6341     6332       -9     
==========================================
- Hits         6338     6330       -8     
+ Misses          3        2       -1     

☔ View full report in Codecov by Harness.
📢 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.