Fix repeated evaluation of fx0 in forward gradient computation

[ChrisRackauckas-Claude]

Summary

• Fixed repeated evaluation of fx0 = f(x) inside the loop in finite_difference_gradient!
• Optimizes function evaluations from 2N to N+1 for forward differences when computing gradients
• Maintains full compatibility with both cached and uncached function values

Problem

As reported in #202, when computing forward differences for gradients, the function f(x) was being evaluated N times inside the loop (once per iteration) in addition to the N evaluations for perturbed inputs, resulting in 2N total function evaluations.

Solution

• Moved the fx0 = f(x) computation outside the loop
• Use conditional assignment: fx0 = typeof(fx) != Nothing ? fx : f(x)
• Simplified the loop logic by eliminating conditional branches
• Applied the same optimization to both real and complex gradient computations

Test plan

• Verified gradient accuracy remains unchanged
• Confirmed function evaluation count is reduced from 2N to N+1
• Ran existing test suite (passed core functionality tests)
• Applied SciMLStyle formatting

Performance Impact

For a vector of length N, this reduces function evaluations by ~50%, providing significant performance improvement for expensive functions.

Fixes #202

🤖 Generated with Claude Code