perf: improve small size fft

[gbotrel]

Description

Improve small size fft runtime (~50%) by avoiding overhead for small sizes (< 1024) .

---

Note

Optimizes small FFT sizes by bypassing recursion for n=512 and n=1024.

• Add kerDIFNP_512/kerDITNP_512 and kerDIFNP_1024/kerDITNP_1024 in field/babybear/fft/fft.go and field/koalabear/fft/fft.go
• Update difFFT/ditFFT to directly call these kernels when stage >= twiddlesStartStage and n ∈ {512,1024}
• Extend generator template internal/generator/field/template/fft/fft.go.tmpl to emit the same kernels and conditional dispatch (guarded by HasASMKernel)

^{Written by Cursor Bugbot for commit 21ae715. This will update automatically on new commits. Configure here.}

[Copilot]

Pull request overview

This PR improves small FFT performance (~50% for sizes < 1024) by adding specialized kernels for 512 and 1024-element FFTs that avoid recursion overhead. The changes introduce new optimized paths that perform outer butterfly stages inline before delegating to existing 256-element kernels.

• Adds dedicated kernels (kerDIFNP_512, kerDITNP_512, kerDIFNP_1024, kerDITNP_1024) for both DIF and DIT FFT algorithms
• Updates dispatch logic to route 512 and 1024-element FFTs to specialized kernels when ASM support is available
• Generated code applied to babybear and koalabear field implementations

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated no comments.

File	Description
internal/generator/field/template/fft/fft.go.tmpl	Adds template code for 512/1024 element kernels and updates dispatch logic with new size checks (conditional on HasASMKernel)
field/koalabear/fft/fft.go	Generated code: adds four new kernel functions and updates difFFT/ditFFT dispatch to check for n==512 and n==1024
field/babybear/fft/fft.go	Generated code: adds four new kernel functions and updates difFFT/ditFFT dispatch to check for n==512 and n==1024