Add FFT related operators and APIs

[iclementine]

PR types

New features

PR changes

OPs

Describe

Add fft related functionalities.

paddle.fft subpackage

1. differentiable 1d to nd complex to complex fast fourier transform. (fft, fft2, fftn, ifft, ifft2, ifftn)
2. differentiable 1d to nd real to complex fast fourier transform. (rfft, rfft2, rfftn, ihfft, ihfft2, ihfftn)
3. differentiable 1d to nd complex to real fast fourier transform. (hfft, hfft2, hfftn, irfft, irfft2, irfftn)
4. differentiable helper functions. (fftfreq, rfftfreq, fftshift, ifftshift)

paddle.signal subpackage

1. short-timw fast fourier transform( stft)
2. inverse short-time fast fourier transform( istft)

Operators

1. fft_c2c, fftr2c, fft_c2r, fft_c2c_grad, fftr2c_grad, fft_c2r_grad with backends on CPU(mkl_cdft and pocket_fft) and GPU(cufft for nvidia gpus and hipfft for amd gpus.) for differentiable fast fourier transforms.
2. frame and overlap_add for stft and istft functionalities.

Others

1. add kernel with complex data type for many operators.
2. add to the functionality for complex data type and transforms between complex and real data types.

Examples

It works like scipy.fft

import numpy as np
import paddle
x = np.random.randn(2, 4) + np.random.randn(2, 4)
xt = paddle.to_tensor(x)
print(np.allclose(np.fft.fftn(x), paddle.fft.fftn(xt).numpy()))

import scipy.fft
import paddle
x = np.random.randn(2, 4)
xt = paddle.to_tensor(x)
print(np.allclose(scipy.fft.rfftn(x), paddle.fft.rfftn(xt).numpy()))

import scipy.fft
import paddle
x = np.random.randn(2, 4) + np.random.randn(2, 4)
xt = paddle.to_tensor(x)
print(np.allclose(scipy.fft.hfftn(x), paddle.fft.hfftn(xt).numpy()))

And more than that.

[chenwhql]

LGTM for PADDLE_THROW(status)，我们的cuda关联第三方库都提供了官方的错误信息作为提示，希望后续能够补充，只报一个错误码还不太友好，可以参考 #33003

[XiaoguangHu01]

LGTM

[Xreki]

LGTM for op benchmark ci

[iclementine]

LGTM for PADDLE_THROW(status)，我们的cuda关联第三方库都提供了官方的错误信息作为提示，希望后续能够补充，只报一个错误码还不太友好，可以参考 #33003

目前的做法是报错误信息的，但是没有加到 proto 里面，我们后续试试把 cufft 和 mkl-cdft 的报错也都加到里面去。