Specialize GEProjDouble with feSquare

[koba-e964]

Benchmark results (Apple M2, go test ./secp256k1 -run=^$ -bench='GEJacobianPoint_|GEProjPoint_' -benchmem -count=5)

Main (origin/main @ 7761398):

• BenchmarkGEJacobianPoint_VariableTime_Short-8: 4206 ns/op
• BenchmarkGEJacobianPoint_VariableTime_Long-8: 1220375 ns/op
• BenchmarkGEJacobianPoint_ConstantTime_Short-8: 1466087 ns/op
• BenchmarkGEJacobianPoint_ConstantTime_Long-8: 1412107 ns/op
• BenchmarkGEProjPoint_ConstantTime_Short-8: 909115 ns/op

geprojdouble-only:

• BenchmarkGEJacobianPoint_VariableTime_Short-8: 5368 ns/op
• BenchmarkGEJacobianPoint_VariableTime_Long-8: 1276801 ns/op
• BenchmarkGEJacobianPoint_ConstantTime_Short-8: 1487202 ns/op
• BenchmarkGEJacobianPoint_ConstantTime_Long-8: 1445619 ns/op
• BenchmarkGEProjPoint_ConstantTime_Short-8: 918919 ns/op

Delta vs main:

• BenchmarkGEJacobianPoint_VariableTime_Short-8: +27.63%
• BenchmarkGEJacobianPoint_VariableTime_Long-8: +4.62%
• BenchmarkGEJacobianPoint_ConstantTime_Short-8: +1.44%
• BenchmarkGEJacobianPoint_ConstantTime_Long-8: +2.37%
• BenchmarkGEProjPoint_ConstantTime_Short-8: +1.08%

Follow-up sanity check (Feb 1, 2026):
go test ./... -bench=GEProjPoint_ConstantTime =>
BenchmarkGEProjPoint_ConstantTime_Short-8: 992828 ns/op (0 B/op, 0 allocs/op)

Note: this PR now also adds a skills note with these benchmark results in skills/issue-8-complete-addition.md.

[koba-e964]

Benchmark results seem fragile. I confirmed that on GitHub Action runners the performance improved somewhat. Merging