[export] fix stft decomp and making it consistent with cpp impl. (pytorch#149232)

Summary: We change the fake impl of stft to follow more closely with its cpp implementation [here](https://github.com/pytorch/pytorch/blob/main/aten/src/ATen/native/SpectralOps.cpp#L951-L963)

where  " n_frames = 1 + (len - n_fft) / hop_length;" is also an integer division.

Test Plan: Existing tests and buck2 build --flagfile fbcode//mode/dev fbcode//executorch/examples/models/fb/llama4:speech_transform.pte

Differential Revision: D71209142

edit: we kept the original path un-changed.

Pull Request resolved: pytorch#149232
Approved by: https://github.com/jackzhxng

Author: ydwu4

torch/_refs/__init__.py

@@ -3451,10 +3451,12 @@ def stft(
         left = (n_fft - win_length_) // 2
         window = aten.constant_pad_nd(window, [left, n_fft - win_length_ - left])
 
-    input = input.unfold(dimension=-1, size=n_fft, step=hop_length_)
     if not center and align_to_window:
         input_pad_amount = (n_fft - win_length_) // 2
         input = aten.pad(input, [input_pad_amount, input_pad_amount], pad_mode)
+
+    input = input.unfold(dimension=-1, size=n_fft, step=hop_length_)
+
     if window is not None:
         input = input * window