[CUFFT] Preallocate a buffer for complex-to-real FFT

[amontoison]

@maleadt
Related to our discussion on Slack:

• https://julialang.slack.com/archives/C689Y34LE/p1733470839241359
  and these issues:
• https://discourse.julialang.org/t/possible-bug-using-plan-irfft-and-plan-rfft-with-mul-and-ldiv/57374
• Inverse Complex-to-Real FFT allocates GPU memory #2249

[maleadt]

Couple of minor comments. CI failures are also related.

[codecov]

Codecov Report

Attention: Patch coverage is 95.65217% with 2 lines in your changes missing coverage. Please review.

Project coverage is 73.69%. Comparing base (ca8f6cf) to head (6f82697).
Report is 1 commits behind head on master.

Files with missing lines	Patch %	Lines
lib/cufft/fft.jl	95.65%	2 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #2578      +/-   ##
==========================================
+ Coverage   73.67%   73.69%   +0.01%     
==========================================
  Files         156      156              
  Lines       15030    15048      +18     
==========================================
+ Hits        11074    11090      +16     
- Misses       3956     3958       +2     

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

[maleadt]

A dedicated typevar instead of simply Union{Nothing,CuArray} seems overkill, but I guess you want to eliminate the partial type information?

[amontoison]

Yes, and I also take this PR as an opportunity to have a concrete type for region in the structure.

[maleadt]

I trust you have good reasons for this, but it's not always advantageous to do so. The CUTENSOR wrappers, for example, cause excessive specialization for little gain, since these calls are very coarse-grained anyway (yet leading to very long test times). In this case, though, the specialization is already there (since B doesn't introduced new information already contained in inplace and T).

[amontoison]

I haven't changed the dispatch for the high-level FFT routines, so I didn't expect any slowdown. I simply made the low-level structure fully inferable.
I'll run a quick regression test for fft/ifft before and after this PR to check if there's any noticeable difference.

I didn't test it without the type B in the structure, but I observed a nice speed-up in my code when adding the buffer inside for rfft/irfft.
Julia also stopped crashing due to OOM issues.